# LLMS-Explorer — facts
> Source-anchored units extracted from the docs: 1769 across 111 pages. Each line ends in the page URL and anchor it came from.
## The API: what it is, and how to call it
- [parameter] `read`: Queries within your quota; read your own artifacts — https://llms-explorer.com/reference/api/#three-ways-in
- [parameter] `run`: Create metered jobs — lint model passes, indexing, concept packs — https://llms-explorer.com/reference/api/#three-ways-in
- [parameter] `publish`: Contribute to the shared catalogue — https://llms-explorer.com/reference/api/#three-ways-in
- [definition] The API: what it is, and how to call it — LLMS-Explorer as a service, not a website: what a key buys you, the REST surface, the hosted MCP gateway, the llmsx CLI, and where the free tier ends. — https://llms-explorer.com/reference/api/#the-api-what-it-is-and-how-to-call-it
- [definition] What you're actually buying — LLMS-Explorer is the concept-family-tree explorer with the hub's llms tooling hung off its nodes: lint or optimize an llms file, turn notes into one, abstract a concept out of a corpus, deepen it with a research wave, map its family. — https://llms-explorer.com/reference/api/#what-youre-actually-buying
- [definition] The free tier's semantic search — The free tier's semantic and hybrid search is a fixed 16-document demo, rate-limited and not billed — it's there to show what the retrieval looks like, not to serve your own corpus. Indexing your own docset, and querying it, both need a paid plan. — https://llms-explorer.com/reference/api/#the-free-tiers-semantic-search
- [definition] What's deterministic and free even with a key — Not every tool call spends a token. Lint's deterministic passes — the ones that don't run a model — are free and unlimited-frequency-within-quota on every plan; only the *model* passes (P4/P8/P12) are metered. — https://llms-explorer.com/reference/api/#whats-deterministic-and-free-even-with-a-key
## The attribute rubric
- [parameter] I1: Attribute=Exactly one H1 naming the site/product (not a page); Applies=index, facts, family; Measure=deterministic; Bar=1 H1; title = product/site; Miss=High — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] I2: Attribute=Blockquote summary immediately after H1, 1–3 sentences, self-contained; Applies=index, family; Measure=deterministic + judgment; Bar=present; says what the thing is and who it is for; Miss=Medium — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] I3: Attribute=Free-form info before the first H2 (how to read this file, versions, languages); Applies=index, family; Measure=judgment; Bar=only if it changes how a reader should use the links; Miss=Low — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] I4: Attribute=Sections are H2 only; each is a link list; no prose after the first H2 except list notes; Applies=index, family; Measure=deterministic; Bar=no H3+, no stray paragraphs; Miss=Medium — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] I5: Attribute=Link entries match `- [name](url)` + optional `: notes`; Applies=index, family; Measure=deterministic; Bar=100% of list items; Miss=High if <90%, else Medium — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] I6: Attribute=Kind is unambiguous from the first 20 lines (index vs full vs facts) — a full file is never served as an index; Applies=all; Measure=deterministic; Bar=grammar detected with one candidate; Miss=High — https://llms-explorer.com/reference/attributes/#1-identity-and-shape
- [parameter] N1: Attribute=Two hops: index → page (or family → index → page); no index links a bare directory of more indexes; Applies=index, family; Measure=deterministic (link targets); Bar=≤2 hops to any page; Miss=High — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N2: Attribute=Section design mirrors how users ask (task/topic groups), not the URL tree or an alphabet; Applies=index; Measure=judgment; Bar=≥80% of sections are task/topic named; Miss=Medium — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N3: Attribute=Ordering by expected query frequency: quickstart/auth/reference/errors first; the first 20% of links should answer 80% of questions; Applies=index; Measure=judgment + agent test; Bar=hot pages in the first section; Miss=Medium — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N4: Attribute=`## Optional` holds only skippable material (changelog, legal, old posts, appendices); it is the last section; Applies=index; Measure=deterministic + judgment; Bar=last; no reference/pricing inside; Miss=Medium — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N5: Attribute=Every page the source publishes that a reader would need is reachable (coverage); Applies=index; Measure=deterministic vs source page list; Bar=≥95% of `reference`+`guide` pages linked; Miss=High if <80% — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N6: Attribute=No dead ends: each link resolves (200, markdown or `.md` twin), no redirect to an HTML app shell; Applies=index, family; Measure=deterministic (`--check-links`); Bar=0 dead links; Miss=High — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] N7: Attribute=Cross-cutting material (errors, auth, glossary) linked once, not once per section; Applies=index, family; Measure=judgment; Bar=no duplicate targets; Miss=Low — https://llms-explorer.com/reference/attributes/#2-navigation
- [parameter] D1: Attribute=Every link carries a description; Applies=index, family; Measure=deterministic; Bar=100%; Miss=Medium (High if <60%) — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] D2: Attribute=Description says what the reader FINDS there, with the exact tokens (flags, env vars, error strings) — not a restated title; Applies=index; Measure=judgment; Bar="Authentication docs." fails; "API key creation, OAuth scopes, token rotation. Required before any call." passes; Miss=Medium — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] D3: Attribute=Length 10–25 words; no trailing ellipsis from truncation; Applies=index; Measure=deterministic; Bar=95% within band; Miss=Low — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] D4: Attribute=No duplicate descriptions across links; Applies=index; Measure=deterministic; Bar=0 duplicates; Miss=Medium — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] D5: Attribute=Descriptions are extractive or verified — model-written ones audited against the page; Applies=index; Measure=judgment (sampled); Bar=sample of 10: 0 hallucinated claims; Miss=High — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] D6: Attribute=Family lines carry counts (pages, ~tokens) so a consumer can budget; Applies=family; Measure=deterministic; Bar=100% of product links; Miss=Medium — https://llms-explorer.com/reference/attributes/#3-descriptions
- [parameter] C1: Attribute=One declared page grammar, stated in a header comment; every page block parses; Applies=full; Measure=deterministic (`split_llms_full`); Bar=blocks parsed = blocks present; Miss=High — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C2: Attribute=Every page has a title and a resolvable source URL; Applies=full; Measure=deterministic; Bar=100%; Miss=High — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C3: Attribute=No navigation residue: "Documentation Index" blockquotes, `[Skip to content]`, MDX wrappers, `theme={null}` props; Applies=full; Measure=deterministic; Bar=0 hits; Miss=Medium — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C4: Attribute=Code fences intact and language-tagged; tables intact; Applies=full; Measure=deterministic (fence balance, table separators); Bar=balanced; ≥90% fences tagged; Miss=Medium — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C5: Attribute=No duplicated pages (same source URL twice) or near-duplicate bodies (e.g. localized copies); Applies=full; Measure=deterministic + embedding; Bar=0 exact dups; near-dups flagged; Miss=Medium — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C6: Attribute=Units are atomic (1–2 sentences), typed from the allowed set, source-anchored; Applies=facts; Measure=deterministic + judgment; Bar=100% typed; 100% anchored; ≥90% atomic; Miss=High for anchors, Medium otherwise — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] C7: Attribute=Units are true to their source span (no generalisation beyond the page); Applies=facts; Measure=judgment (sampled re-read); Bar=sample of 20: ≥95% supported; Miss=High — https://llms-explorer.com/reference/attributes/#4-content-fidelity-full-and-facts
- [parameter] P1: Attribute=Provenance banner: who generated it, from what, when (`verified-as-of` / `generated` date); Applies=all; Measure=deterministic; Bar=present; Miss=Medium — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] P2: Attribute=Links point at the publisher's canonical URLs (or its `.md` twins), never at a private mirror, unless the file is explicitly internal; Applies=index; Measure=deterministic; Bar=100% public or file marked internal; Miss=High — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] P3: Attribute=Rights: a third-party `llms-full.txt` is marked internal/private; the index is what is published; Applies=full; Measure=judgment; Bar=marker present when third-party; Miss=High — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] P4: Attribute=No instructions to the reading model ("ignore…", "you must…", "always answer…") — 42% of files in the wild try to steer; ours never do; Applies=all; Measure=deterministic (pattern) + judgment; Bar=0 imperative-to-model spans; Miss=High — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] P5: Attribute=No secrets, tokens, emails, internal hostnames in copied text; Applies=all; Measure=deterministic (patterns); Bar=0 hits; Miss=High — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] P6: Attribute=Volatile claims stamped (versions, prices, "current"); Applies=facts; Measure=judgment; Bar=stamped or dated; Miss=Low — https://llms-explorer.com/reference/attributes/#5-provenance-and-trust
- [parameter] S1: Attribute=Index size ≤ ~10 KB / ~2.5k tokens; over that, split hub-and-spoke (never drop pages); Applies=index; Measure=deterministic; Bar=≤10 KB or split; Miss=Medium (High >100 KB) — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] S2: Attribute=Full file has a size ladder beside it (index, small ≤ ~50k tokens, full) with token counts published; Applies=full; Measure=deterministic (manifest); Bar=small + counts present; Miss=Medium — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] S3: Attribute=Small variant = reference-class pages first, within budget; Applies=small; Measure=deterministic; Bar=≤50k tokens; classes honoured; Miss=Medium — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] S4: Attribute=Facts file ≤ ~15% of the cleaned source prose (compression); Applies=facts; Measure=deterministic; Bar=ratio ≤0.15; Miss=Low (Medium >0.3) — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] S5: Attribute=Token estimate declared with its estimator (chars/4 etc.); Applies=manifest; Measure=deterministic; Bar=present; Miss=Low — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] S6: Attribute=No single page block > 200 KB without a note (changelogs); Applies=full; Measure=deterministic; Bar=flagged; Miss=Low — https://llms-explorer.com/reference/attributes/#6-size-and-budget
- [parameter] R1: Attribute=Keyword index exists for the facts/full text (FTS5 over units/chunks) and returns the exact-token queries (`CLAUDE_CODE_SYNC_SKILLS`, `--append-system-prompt`); Applies=facts, full; Measure=measured; Bar=10/10 exact-token probes hit; Miss=High — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R2: Attribute=Vector index exists (`__facts` collection) and the facts layer answers the golden questions better than raw; Applies=facts; Measure=measured (`query --layer`); Bar=golden score ≥ raw score; Miss=Medium — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R3: Attribute=Anchors are stable (`#slug` of the heading) so a hit can be opened at the span; Applies=facts, full; Measure=deterministic; Bar=100% anchors resolve to a heading; Miss=Medium — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R4: Attribute=Unit text carries the exact tokens in `keywords` so BM25 can find them; Applies=facts; Measure=deterministic; Bar=≥80% of units with a code/flag/env token have it in keywords; Miss=Medium — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R5: Attribute=Agent test: an agent given ONLY the index answers N seeded questions by following ≤2 links; Applies=index; Measure=live agent test; Bar=≥8/10; Miss=High if <6/10 — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R6: Attribute=Facts test: an agent given ONLY the facts file answers the same questions without opening pages; Applies=facts; Measure=live agent test; Bar=≥7/10; Miss=Medium — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] R7: Attribute=Every page in the index has ≥1 unit in the facts file (no silent gaps); Applies=index+facts; Measure=deterministic; Bar=≥95% pages covered; Miss=Medium — https://llms-explorer.com/reference/attributes/#7-retrieval-readiness
- [parameter] F1: Attribute=Family file links indexes, never pages; Applies=family; Measure=deterministic; Bar=100% targets are `llms.txt` files; Miss=High — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] F2: Attribute=Each product line carries page + token counts and, where present, a facts link; Applies=family; Measure=deterministic; Bar=100%; Miss=Medium — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] F3: Attribute=Shared material (errors, auth, glossary) appears once, in the family file; Applies=family; Measure=judgment; Bar=no duplication into products; Miss=Low — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] F4: Attribute=The most-specific rule holds: a product's own index is authoritative for its pages; the family never restates them; Applies=family; Measure=judgment; Bar=no page links; Miss=Medium — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] F5: Attribute=Family membership matches the concept tree / hub taxonomy it claims to represent; Applies=family; Measure=deterministic vs tree; Bar=100% of tree children present; Miss=Medium — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] F6: Attribute=Root → family → product is discoverable by `Link: rel=describedby` from any file; Applies=family; Measure=deterministic (headers); Bar=header present; Miss=Low — https://llms-explorer.com/reference/attributes/#8-family-nesting
- [parameter] H1: Attribute=UTF-8, LF, no tabs in list lines, no trailing whitespace, single trailing newline; Applies=all; Measure=deterministic; Bar=clean; Miss=Hygiene (Low) — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H2: Attribute=`Content-Type: text/markdown; charset=utf-8` (or `text/plain`), HTTP 200, no redirect, no auth on the path; Applies=served; Measure=deterministic (HEAD); Bar=pass; Miss=High — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H3: Attribute=`Link: rel=describedby` on files; `rel=alternate type=text/markdown` on HTML pages; Applies=served; Measure=deterministic; Bar=present; Miss=Low — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H4: Attribute=`X-Markdown-Tokens` (or manifest tokens) available before fetch; Applies=served; Measure=deterministic; Bar=present; Miss=Low — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H5: Attribute=Regenerated by the build, not hand-maintained; a `generated` stamp newer than the source; Applies=all; Measure=deterministic (mtime/stamp); Bar=stamp ≥ source mtime; Miss=Medium — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H6: Attribute=Validator-clean on the community validators' strict rules where they do not contradict the spec; Applies=index; Measure=deterministic; Bar=0 High; Miss=Low — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H7: Attribute=Lighthouse agentic audit would not flag it (no 5xx on fetch); Applies=served; Measure=deterministic; Bar=200; Miss=Medium — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] H8: Attribute=`manifest.json` present and consistent with the files (bytes, tokens, pages, units); Applies=export dir; Measure=deterministic; Bar=consistent; Miss=Medium — https://llms-explorer.com/reference/attributes/#9-hygiene-and-serving
- [parameter] Purpose: index (`llms.txt`)=orientation + navigation; full (`llms-full.txt`)=whole text in one fetch; facts (`llms-facts.txt`)=the checkable claims, each anchored — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [parameter] Reader: index (`llms.txt`)=an agent deciding where to look; full (`llms-full.txt`)=a big-context agent or an indexer; facts (`llms-facts.txt`)=a retriever answering a question — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [parameter] Unit: index (`llms.txt`)=link + description; full (`llms-full.txt`)=page block; facts (`llms-facts.txt`)=typed unit with source + anchor — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [parameter] Size: index (`llms.txt`)=≤10 KB; full (`llms-full.txt`)=unbounded (ladder beside it); facts (`llms-facts.txt`)=≤15% of prose — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [parameter] Judged mostly on: index (`llms.txt`)=N*, D*; full (`llms-full.txt`)=C1–C5, S*; facts (`llms-facts.txt`)=C6–C7, R*, P4 — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [parameter] Tested by: index (`llms.txt`)=agent test (R5); full (`llms-full.txt`)=grammar round-trip (C1); facts (`llms-facts.txt`)=keyword + vector probes (R1–R2), facts test (R6) — https://llms-explorer.com/reference/attributes/#10-the-three-kinds-side-by-side
- [definition] The attribute rubric — Every attribute an llms file is judged on, with bars and severities. — https://llms-explorer.com/reference/attributes/#the-attribute-rubric
## Changelog: spec v1 to v2, and the hub pipeline
- [parameter] Required elements: v1=H1 + blockquote + sections implied; v2=**H1 only** required; blockquote, prose and sections optional; Effect on an existing file=none required; the lint still scores a missing blockquote as Medium (I2) — quality, not validity — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Placement: v1=`/llms.txt` at the site root; v2=root **or any subpath**; a file covers the URLs under its path; **most-specific wins**; `/.well-known/` explicitly rejected; Effect on an existing file=enables families and split roots (`/llms.txt`) — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Discovery: v1=none; v2=`Link: <…>; rel="describedby"` on files; `rel="alternate" type="text/markdown"` on HTML pages; as `` or an HTTP header; Effect on an existing file=add the two headers ([usage](/reference/usage/) §1) — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Markdown twins: v1=`page.html.md`; v2=`page.html.md` **or** `page.md`; directories append `index.html.md` or `index.md`; Effect on an existing file=either form passes the twin probe (N6) — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] `## Optional`: v1=mechanical: skippable when context is short; consumed by `llms_txt2ctx`; v2=a **convention** for secondary information; `llms_txt2ctx` and context-expansion removed from the proposal; Effect on an existing file=keep it last; build nothing that depends on it — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] BOM: v1=—; v2=an optional byte-order mark is tolerated; Effect on an existing file=the lint strips it as hygiene (P14) — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Consumption model: v1=expand the whole file into context; v2="view or search the index, then follow links"; the index stays small; detail lives behind links; Effect on an existing file=the size ladder (small / full) becomes the producer's job — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Authoring guidance: v1=—; v2=concise language, informative link descriptions, no unexplained jargon, "test your file by asking an agent questions … giving it only your llms.txt"; Effect on an existing file=the agent test (R5, P12) is the spec's own test made numeric — https://llms-explorer.com/reference/changelog/#the-spec-v1-v2-2026-08-10
- [parameter] Acquire: V1 (to 2026-08-29)=trafilatura BFS crawl → banner mirror; V2 (from 2026-08-30)=the ladder: `llms-full.txt` → `llms.txt` + `.md` twins → `Accept: text/markdown` → docs API → structured crawl; the banner mirror stays the internal format — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Clean: V1 (to 2026-08-29)=none (raw HTML → text); V2 (from 2026-08-30)=`docset_refine clean`: boilerplate lines, MDX → markdown, page classes (reference / guide / changelog / marketing / index) — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Extract: V1 (to 2026-08-29)=`distill_offline.py bulk` — zero-LLM, output never consumed; V2 (from 2026-08-30)=`extract` (snippets, table rows → parameter, definitions, changelog → change; anchors to real headings) + `units` (local LLM, evidence rule) + `polish` — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Export: V1 (to 2026-08-29)=none; V2 (from 2026-08-30)=`export_llms`: index (split over 10 KB) / full (Mintlify grammar) / small (≤ 200,000 chars) / facts / manifest; `topical`; `vocabulary` — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Index: V1 (to 2026-08-29)=one raw vector layer; V2 (from 2026-08-30)=raw **and** facts vector layers, plus an FTS5 keyword layer per layer — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Serve: V1 (to 2026-08-29)=`web-text-mirror --serve` (HTML); V2 (from 2026-08-30)=`llms_serve.py`: `/llms.txt`, `/d//…`, `/m//…`, `/t//…`, with the markdown headers — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Gate: V1 (to 2026-08-29)=none; V2 (from 2026-08-30)=`llms_lint.py` (the deterministic passes) inside `docset_rollout cleanup`; `/ldo` for the model and live passes — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] Artifacts: V1 (to 2026-08-29)=`.pages/`, `_master.md`, `._distill_index.json`; V2 (from 2026-08-30)=`.reference/{pages.json, structured.jsonl, units.jsonl, all_units.jsonl}`, `.llms/` — https://llms-explorer.com/reference/changelog/#the-hub-pipeline-v1-v2-2026-08-30
- [parameter] 2026-08-30: `docset_refine` gains `clean / extract / units / polish / render / export`; the reference dir layout above — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-30: `export_llms` writes the four-file ladder plus `manifest.json`; index split at 10,000 bytes; `PART_PAGES = 60` — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-30: `llms_lint.py` ships the deterministic passes P0–P3, P5–P7, P9 and P14 and the `--json` CI output; `UNIT_RE` fixes the facts line grammar — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-30: `llms_serve.py` sends `Content-Type: text/markdown`, `X-Markdown-Tokens`, `Link: rel="describedby"` — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-30: `docset_refine topical` and `vocabulary`; tree nodes carry `slug` / `aliases`; `--register` writes `llmsFile` on a node — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-31: `export_llms` honours `manifest.json["overrides"]` (`title`, `summary`, `section_order`, `note`) so hand inputs survive regeneration — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [parameter] 2026-08-31: `llms_lint.py --kind vocabulary` lints the vocabulary line grammar — https://llms-explorer.com/reference/changelog/#dated-changes-to-the-hub-schema
- [definition] Changelog: spec v1 to v2, and the hub pipeline — What changed in the llms.txt proposal on 2026-08-10, rule by rule, with the effect on an existing file — and the dated changes to the hub schema behind this site. — https://llms-explorer.com/reference/changelog/#changelog-spec-v1-to-v2-and-the-hub-pipeline
## The concept tree: nodes, frontier, and how to read a node page
- [parameter] `concept`: the node's name, and the string its parent and children link it by — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `slug`: its URL segment; stable, and the key the API in step 3 will use — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `parent`: the concept it hangs from — linked, unless the node is a root — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `children`: the concepts it names, each either researched (linked) or frontier (greyed) — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `aliases`: other names the same concept goes by; the filter matches these too — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `researchedAt`: the date the research run that created the node finished — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `sourcesCount`: how many sources that run read — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `conceptsCount`: how many concepts that run identified under this one — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `skillId`: the skill the research produced, when it produced one — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [parameter] `state`: `researched` for every node with a page; `frontier` only for a named child without one — https://llms-explorer.com/reference/concept-tree/#the-fields-on-a-node-page
- [definition] The concept tree: nodes, frontier, and how to read a node page — What the tree is, why frontier is derived rather than stored, what every field on a node means, and how the browser at /tree/ filters it. — https://llms-explorer.com/reference/concept-tree/#the-concept-tree-nodes-frontier-and-how-to-read-a-node-page
- [definition] What a node is — A node is one researched concept. The tree is stored as a **flat list of nodes linked by name** — each node names its parent and its children as strings, not as pointers — so a rename is a one-place edit and a reader can hold the whole file in mind. — https://llms-explorer.com/reference/concept-tree/#what-a-node-is
- [definition] Frontier is derived, never stored — A **frontier** concept on this site is a name that appears in some node's `childConcepts` and has no node of its own. It is computed on every build from the two sides of that comparison, never stored as a status: a stored status can disagree with the tree, and a derived one cannot. — https://llms-explorer.com/reference/concept-tree/#frontier-is-derived-never-stored
- [definition] How the filter works — The filter box on `/tree/` matches a substring against each concept **and its aliases**, and a branch survives if it or any descendant matches — so filtering hides non-matching branches without ever hiding the path to a hit. — https://llms-explorer.com/reference/concept-tree/#how-the-filter-works
- [definition] What is not here yet — Queueing a frontier concept for research, forking the tree, and attaching your own files to a node are per-user actions, and this site has no accounts yet. — https://llms-explorer.com/reference/concept-tree/#what-is-not-here-yet
## Concept facts files
- [definition] Concept facts files — Load the preserved concept tree as source-cited Markdown facts. — https://llms-explorer.com/reference/context-files/#concept-facts-files
- [definition] Read a concept — Fetch `/context.md` to choose a concept, then fetch its `/downloads/concepts/.md` file. Each file includes its summary, facets, facts and related concepts. — https://llms-explorer.com/reference/context-files/#read-a-concept
## Ethos: what an llms file owes its reader
- [parameter] An index — links and extractive descriptions: yes; it is a map of someone's public pages — https://llms-explorer.com/reference/ethos/#5-rights-are-explicit
- [parameter] A facts file — short anchored claims, each traceable: yes; quotation with attribution, bounded in length — https://llms-explorer.com/reference/ethos/#5-rights-are-explicit
- [parameter] Your own words — hand pages, essays, this site: yes — https://llms-explorer.com/reference/ethos/#5-rights-are-explicit
- [parameter] Third-party full text — a mirrored `llms-full.txt` of a site you do not own: served only to its owner, or under the internal marker; never on a public route — https://llms-explorer.com/reference/ethos/#5-rights-are-explicit
- [definition] Ethos: what an llms file owes its reader — Files are promises; generate, do not hand-edit; never instruct the reader; evidence is external; rights are explicit. — https://llms-explorer.com/reference/ethos/#ethos-what-an-llms-file-owes-its-reader
- [definition] 1. Files are promises — Every link resolves. Every fact is anchored to a heading that exists. — https://llms-explorer.com/reference/ethos/#1-files-are-promises
- [definition] 2. Generate, don't hand-edit — An llms file is an output. Its inputs are a mirror, a page list, extracted units, a concept tree, and a small overrides file (`title`, `summary`, `section_order`, `note`). — https://llms-explorer.com/reference/ethos/#2-generate-dont-hand-edit
- [definition] 3. Never instruct the reader — A docs file has no business telling a model what to say. The spec repository's issue #152 found 42.3% of a sample of wild files attempting exactly that ([evidence](/reference/evidence/)). — https://llms-explorer.com/reference/ethos/#3-never-instruct-the-reader
- [definition] 4. Evidence is external — A finding about an llms file cites something outside the file: the link check, the mirror span behind a unit, the probe result, the HTTP response. A finding with no external evidence is Low at most. — https://llms-explorer.com/reference/ethos/#4-evidence-is-external
- [definition] 5. Rights are explicit — Three tiers, and the tooling knows which is which: — https://llms-explorer.com/reference/ethos/#5-rights-are-explicit
- [definition] The test — If a stranger's agent, handed only the index, can answer eight of ten reasonable questions in two hops — below six is a High — and can check any facts line it relies on in one fetch, the file kept its promises. Nothing else on this site is a stronger claim than that. — https://llms-explorer.com/reference/ethos/#the-test
## Ecosystem evidence
- [parameter] Feb→May 2025: Source=Chris Green; Sample=Majestic Million; Finding=15 → 105 valid files (~0.01%); ~100k crawl errors caveat[^4] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Jun 2025: Source=Originality.ai; Sample=3M+ sites; Finding=4,088 llms.txt[^5] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Jun 2025: Source=Rankability; Sample=Tranco top 1,000; Finding=0.3%[^6] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Jul 2025: Source=HTTP Archive (Burridge); Sample=top 10k; Finding=1.04% valid[^7] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Nov 2025: Source=SE Ranking; Sample=~300k domains; Finding=10.13% overall (9.88% low-traffic / 10.54% mid / 8.27% 100k+ visits)[^3] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Mar 2026: Source=Originality.ai via ppc.land; Sample=Fortune 500; Finding=7.4% (37/500)[^8] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] May 2026: Source=Originality.ai; Sample=3M+ sites; Finding=36,120 llms.txt (8.8× YoY); llms-full.txt 23 → 2,463 (107×); ai.txt 397[^5] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] May 2026: Source=Ahrefs; Sample=137,210 Ahrefs-Web-Analytics domains; Finding=28% publish a valid file (self-selected sample)[^1] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Jun 2026: Source=HTTP Archive (Burridge); Sample=top 1k / 10k / 100k / 1M; Finding=6.28% / 5.61% / 5.17% / 5.07% (~5.4× in 12 months)[^7] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Jun 2026: Source=Rankability; Sample=Tranco top 1,000; Finding=8.7% (87 files; 15 with llms-full.txt)[^6] — https://llms-explorer.com/reference/evidence/#2-adoption-measurements-dated
- [parameter] Ahrefs (2026-06-15): Window / sample=May 2026 logs, 137,210 domains; Finding=**97% of valid files got zero requests**; of requests, 96% bots, 77% of those non-AI (SEO auditors 21.7%); named AI bots 19.5%; AI training crawlers 5.3% (GPTBot 4.51%, ClaudeBot 0.8%); AI retrieval 1.1% (OAI-SearchBot 0.74%); **0 AI requests to non-existent files** (nobody probes speculatively); the `Claude-Code` UA… — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] OtterlyAI (2026-02-05): Window / sample=90 days, one site; Finding=84 of 62,100 AI-bot requests hit /llms.txt (0.1%)[^15] — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] Wislr (Feb–Mar 2026): Window / sample=48 days, one site; Finding=12,099 bot requests; robots.txt fetched hundreds of times (OAI-SearchBot 180, ClaudeBot 175); sitemap.xml too; **llms.txt 0**[^16] — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] EZY Research (Apr–Jul 2026): Window / sample=83 sites, 12 weeks; Finding=robots vs llms: GPTBot 3,990/7, ClaudeBot 3,120/9, PerplexityBot 775/0, Googlebot 5,125/67, **Meta-ExternalAgent 172/193** (the only bot fetching it more than robots.txt)[^17] — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] Hacker News thread (Feb 2026): Window / sample=anecdotal logs; Finding=only OVH/GCP-hosted tools (WebPageTest, BuiltWith), no ChatGPT/Claude UAs[^18] — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] Cloudflare `Accept: text/markdown` (Mar–Apr 2026): Window / sample=44 days, one Worker; Finding=1,421 requests: headless Chrome 639, "Claude" (Anthropic infra) 500, axios 211; no GPTBot/PerplexityBot/ClaudeBot[^19] — https://llms-explorer.com/reference/evidence/#4-who-reads-server-log-studies
- [parameter] directory.llmstxt.cloud: Size="4,000 websites listed" (49M llms.txt tokens / 325M llms-full tokens); Notes=named in spec v2 — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
- [parameter] llmstxthub.com: Size=~2,650 entries, 15–16 categories (David Dias); Notes=named in spec v2 — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
- [parameter] llmstxt.site: Size=~1,000+ (≈170 in May 2025); columns product / website / llms.txt / llms-full.txt / **token counts**; `/submit`; Notes=named in spec v2 — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
- [parameter] SecretiveShell/Awesome-llms-txt: Size=784 link lines (counted 2026-08-30); Notes=GitHub — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
- [parameter] llms-text.com: Size="780+ verified implementations"; Notes=vendor's own directory — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
- [parameter] llms-text.com/blog/sites-using-llms-txt: Author / date=Michael Vereb, 2025-07-25; Claims="780+ verified"; names Anthropic, Cloudflare, Supabase, Vercel, ElevenLabs, Firecrawl, Mintlify, Cursor, Aptos, GitBook, Wix; "no e-commerce adoption"; Grade=adopters check out on live probe; count uncorroborated — low for numbers, fine for examples[^31] — https://llms-explorer.com/reference/evidence/#7-vendor-sources-graded
- [parameter] llms-text.com/blog/what-is-llms-txt: Author / date=same; Claims="foundational pillar of GEO"; ChatGPT/Perplexity/Cursor/Windsurf/Claude Code consume it; "up to 114% more tokens" (incoherent arithmetic), "10–15% accuracy" — unattributed; Grade=GEO and ChatGPT/Perplexity-consumption claims contradicted by every log study — low[^32] — https://llms-explorer.com/reference/evidence/#7-vendor-sources-graded
- [parameter] llms-text.com/blog/llms-txt, /how-to-create-llms-txt: Author / date=same; Claims=MIME/200/UTF-8 rules; `Link: …; rel="describedby"`; "under 10 KB"; framework snippets; funnels to its generator/validator; Grade=useful mechanics (the `describedby` relation is now in spec v2), vendor numbers — medium[^33][^34] — https://llms-explorer.com/reference/evidence/#7-vendor-sources-graded
- [parameter] gitdoc.ai/blog/llms-txt-ai-readable-documentation: Author / date=Yadian Llada / GitDoc, 2026-05-22; Claims=GitBook: 41% of docs page requests from AI agents (unverified); permission / inventory / navigation distinction; curate 10–20 pages (quickstart, auth, per-resource reference, errors, changelog); regenerate in the build; llms-full for priority pages; Grade=sound guidance, unverified headline… — https://llms-explorer.com/reference/evidence/#7-vendor-sources-graded
- [definition] 1. The one-line verdict — Adoption is real and growing (≈5–10% of the general web by mid-2026, 28% among SEO-savvy sites, 8.8× year on year); *unsolicited* consumption is near zero (97% of files never get an AI request); the demonstrated use is agents that are pointed at the file — the `Claude-Code` UA out-fetched every AI… — https://llms-explorer.com/reference/evidence/#1-the-one-line-verdict
- [definition] 6. Directories and registries — Self-submitted, overlapping, unverified — lower bounds, not measurements:[^27][^28][^29][^30] — https://llms-explorer.com/reference/evidence/#6-directories-and-registries
## Formatting: the grammars side by side
- [snippet] 1. The index — `llms.txt`: > One paragraph saying what this is and who it is for. — `# Product` — https://llms-explorer.com/reference/formatting/#1-the-index-llmstxt
- [snippet] 4. The facts line — `llms-facts.txt`: https://example.com/docs/install.md — `## Install` — https://llms-explorer.com/reference/formatting/#4-the-facts-line-llms-factstxt
- [snippet] 5. The vocabulary line — `llms-vocabulary.txt`: - **anchor** [llms.anchor] (noun): the `#fragment` on a facts-line URL that names the heading a claim came from — https: — `- **anchor** [llms.anchor] (noun): the `#fragment` on a facts-line URL that names the heading a clai` — https://llms-explorer.com/reference/formatting/#5-the-vocabulary-line-llms-vocabularytxt
- [parameter] Mintlify: Page block=`# Title` / `Source: ` / blank / body; blank lines between pages; Who=Mintlify sites, Claude Code docs, **the hub** (`GRAMMAR_NOTE`) — https://llms-explorer.com/reference/formatting/#2-the-full-file-llms-fulltxt
- [parameter] Anthropic YAML: Page block=site H1, `---`, per page `## Heading` + YAML (`title:` / `url:` / `description:`) + raw MDX; Who=platform.claude.com — https://llms-explorer.com/reference/formatting/#2-the-full-file-llms-fulltxt
- [parameter] Cloudflare frontmatter: Page block=YAML frontmatter, a "Documentation Index" blockquote, `# Title`, `[View as Markdown](…/index.md)`, body; Who=developers.cloudflare.com — https://llms-explorer.com/reference/formatting/#2-the-full-file-llms-fulltxt
- [definition] Formatting: the grammars side by side — The index, the three full-file grammars, the facts line, the small file, the vocabulary line, the split root and the manifest — on one page. — https://llms-explorer.com/reference/formatting/#formatting-the-grammars-side-by-side
- [definition] 1. The index — `llms.txt` — The only file the spec defines. Structure, in order: an optional BOM; **one H1** naming the site or product (the only required element); a blockquote summary of one to three sentences; free-form prose (no headings) about how to read the file; then H2 sections, each a list of links. — https://llms-explorer.com/reference/formatting/#1-the-index-llmstxt
- [definition] 2. The full file — `llms-full.txt` — Not in the spec; three grammars are in the wild. The hub emits the first and names it in a header comment so a parser never has to guess. — https://llms-explorer.com/reference/formatting/#2-the-full-file-llms-fulltxt
- [definition] 3. The budgeted file — `llms-small.txt` — Same grammar as the full file, different selection: reference-class pages first, then guides, until `SMALL_MAX_CHARS = 200_000` characters (about 50k tokens at `CHARS_PER_TOKEN = 4`) — the ceiling at which indexed docs become unstable in consumers such as Cursor. — https://llms-explorer.com/reference/formatting/#3-the-budgeted-file-llms-smalltxt
- [definition] 4. The facts line — `llms-facts.txt` — A hub extension: the checkable claims, one per line, each anchored to the heading it came from. — https://llms-explorer.com/reference/formatting/#4-the-facts-line-llms-factstxt
- [definition] 5. The vocabulary line — `llms-vocabulary.txt` — The lexical layer, spec-v2-shaped so any llms reader can open it: an H1 ` — vocabulary`, a blockquote with the term count, then `## Terms`, `## Homonyms` and `## Named, not yet defined`. One line per term per sense: — https://llms-explorer.com/reference/formatting/#5-the-vocabulary-line-llms-vocabularytxt
- [definition] 6. Split roots and families — When an index would exceed 10 KB the sections become subpath indexes: the root keeps the H1, blockquote and a `## Sections` list of `/llms.txt` links, each line carrying page and token counts; a section with no further path structure is cut into `part-N` files of `PART_PAGES = 60` pages. — https://llms-explorer.com/reference/formatting/#6-split-roots-and-families
- [definition] 7. The manifest — `manifest.json` — Beside the files, never linked from them: `files{name: {bytes, tokens}}`, `chars_per_token`, `pages`, `units`, `sections`, `dropped_empty_pages`, `acquired` (how the mirror was obtained), and `overrides` — the hand inputs (`title`, `summary`, `section_order`, `note`) that survive regeneration. — https://llms-explorer.com/reference/formatting/#7-the-manifest-manifestjson
- [definition] Reading order — Index first, always. Fall through to `llms-small.txt` when you need whole pages and have a budget, to `llms-full.txt` when you have none, to `llms-facts.txt` when you need a claim with a place to check it. — https://llms-explorer.com/reference/formatting/#reading-order
## Glossary
- [definition] Glossary — The terms of the field, one line each, in the sense this site uses them — with the contrasts that matter. — https://llms-explorer.com/reference/glossary/#glossary
## The passes
- [parameter] B0: Passes=P0; Kind=deterministic; Runs as=inline, first — every other pass keys off the detected kind — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B1: Passes=P1 P2 P3 P5 P14; Kind=deterministic; Runs as=one `llms_lint.py` invocation, JSON findings — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B2: Passes=P4 P9; Kind=model; Runs as=one subagent reading the index (+ sample pages) — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B3: Passes=P6 P7; Kind=deterministic; Runs as=`llms_lint.py --full` / `--facts` (same invocation as B1 when the kind is full/facts) — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B4: Passes=P8; Kind=model, sampled; Runs as=one subagent, 20 units re-read against source spans — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B5: Passes=P10; Kind=deterministic + model; Runs as=only when kind = family or `--family` — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B6: Passes=P11; Kind=live; Runs as=`docset_indexer.py keyword` + `query --layer facts` probes — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B7: Passes=P12; Kind=live agent; Runs as=one fresh-context subagent given ONLY the file; opt-in `--agent-test`, default on for new files — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B8: Passes=P13; Kind=live HTTP; Runs as=only when a URL is given or `--serve-check` — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [parameter] B9: Passes=P15; Kind=deterministic; Runs as=only when the export directory has a source mirror — https://llms-explorer.com/reference/passes/#bundle-map-and-dispatch-rules
- [definition] The passes — What the optimizer runs, in order, and how each pass is judged and fixed. — https://llms-explorer.com/reference/passes/#the-passes
- [definition] Severity resolution across passes — Same span flagged by several passes: take the highest severity; tie → lower pass number wins (P0 > P1 > …); tie → the more conservative fix (report over rewrite). — https://llms-explorer.com/reference/passes/#severity-resolution-across-passes
- [definition] N/A rules — A pass reports `N/A ()` — never silently skips — when: the kind excludes it (P6 on an index), the layer is absent (P11 without an indexed docset), the input is missing (P13 without a URL, P15 without a mirror), or the flag is off (P12 without `--agent-test` on a refresh run). — https://llms-explorer.com/reference/passes/#na-rules
## Reasoning: why the rules are what they are
- [definition] Reasoning: why the rules are what they are — Extractive descriptions, the size ladder, anchors, facts as the trusted layer, and the two-hop bar — each rule traced to the evidence that produced it. — https://llms-explorer.com/reference/reasoning/#reasoning-why-the-rules-are-what-they-are
- [definition] 1. Extractive descriptions beat generated ones — An index description exists so that a routing model can decide, without fetching, whether the page answers its question. That decision is made on tokens: the flag name, the error string, the endpoint path. — https://llms-explorer.com/reference/reasoning/#1-extractive-descriptions-beat-generated-ones
- [definition] 2. Size is a producer-side problem — Consumers do not truncate gracefully. — https://llms-explorer.com/reference/reasoning/#2-size-is-a-producer-side-problem
- [definition] 3. Anchors make facts checkable — A claim without a place to verify it is a rumour with a URL. The facts line carries `url#anchor`, and the anchor must resolve to a heading that exists on the page (attribute C6). — https://llms-explorer.com/reference/reasoning/#3-anchors-make-facts-checkable
- [definition] 4. The facts file is the trusted layer — Raw page text is untrusted input — the spec repository's own issue #152 found that 42.3% of a 100-file sample tried to steer the reader ([evidence](/reference/evidence/)). — https://llms-explorer.com/reference/reasoning/#4-the-facts-file-is-the-trusted-layer
- [definition] 5. Two hops, from the index alone — The spec's own test is the bar: "test your file by asking an agent questions about your content, giving it only your llms.txt as a starting point." The rubric makes it numeric — attribute R5, pass P12: ten questions, an agent that starts from the index and may follow links, at least eight answered… — https://llms-explorer.com/reference/reasoning/#5-two-hops-from-the-index-alone
- [definition] 6. Publish for agents, not for search — The evidence page holds the numbers: adoption at roughly 5–10% of the general web by mid-2026 and rising 8.8× year on year, yet 97% of valid files received zero AI requests in a month of logs, Google says it does not read the file, and a 300k-domain model found no relationship between having one… — https://llms-explorer.com/reference/reasoning/#6-publish-for-agents-not-for-search
- [definition] 7. Why the rubric is deterministic first — Every attribute names its measure: deterministic, model judgment, or live. The lint implements the deterministic passes P0–P3, P5–P7, P9 and P14 with no model call and gates CI on them; the model, live and family passes (P4, P8, P10–P13, P15) run under `/ldo` when someone asks. — https://llms-explorer.com/reference/reasoning/#7-why-the-rubric-is-deterministic-first
## Recreating and aggregating
- [snippet] 6. Scale to a family: nested indexes, hub-and-spoke: > One index per product below; each product's own llms.txt is the authoritative map of that product. — `# Acme Platform docs` — https://llms-explorer.com/reference/recreation/#6-scale-to-a-family-nested-indexes-hub-and-spoke
- [definition] Recreating and aggregating — The acquisition ladder, lenient parsing, families, rights. — https://llms-explorer.com/reference/recreation/#recreating-and-aggregating
- [definition] 2. Acquire clean markdown — the ladder — Try in this order; each step is cheaper and cleaner than the next: 1. — https://llms-explorer.com/reference/recreation/#2-acquire-clean-markdown-the-ladder
- [definition] 3. Build the index for ONE product — Structure (spec v2): H1 = product name; blockquote = one-paragraph summary; optional prose "how to interpret the files"; H2 sections, each a list of `- [name](url): description`.[^4] — https://llms-explorer.com/reference/recreation/#3-build-the-index-for-one-product
- [definition] 6. Scale to a family: nested indexes, hub-and-spoke — Spec v2 gives the mechanism: "The file can be placed at the site root, or at any path within it, covering the pages under that path … where more than one file applies, agents should use the most specific one."[^4] The live exemplar is **Cloudflare**: `developers.cloudflare.com/llms.txt` holds ~105… — https://llms-explorer.com/reference/recreation/#6-scale-to-a-family-nested-indexes-hub-and-spoke
## llms.txt: the spec and its grammars
- [snippet] 2. The spec, v2: > Optional description goes here — `# Title` — https://llms-explorer.com/reference/spec/#2-the-spec-v2
- [parameter] Spec-conformant: Example=code.claude.com/docs/llms.txt, FastHTML; Notes=H1, blockquote, H2 sections, `.md` links — https://llms-explorer.com/reference/spec/#31-llmstxt-three-real-shapes
- [parameter] API-first: Example=docs.github.com/llms.txt; Notes=first H2 "How to use" lists JSON/markdown APIs (Page List, Article Body → markdown, Search) and the MCP server before any content links[^6] — https://llms-explorer.com/reference/spec/#31-llmstxt-three-real-shapes
- [parameter] Non-conformant prose: Example=docs.anthropic.com/llms.txt; Notes=H1, then prose and `## Root URL` / language lists, no blockquote[^6] — https://llms-explorer.com/reference/spec/#31-llmstxt-three-real-shapes
- [parameter] Mintlify: Page block=`# Title` / `Source: ` / blank / description / body; pages separated by blank lines only; Verified sample=code.claude.com/docs/llms-full.txt (191 pages, 8.5 MB); mintlify.com/docs/llms-full.txt[^7] — https://llms-explorer.com/reference/spec/#32-llms-fulltxt-not-in-the-spec-and-three-grammars
- [parameter] Anthropic platform: Page block=site H1, `---`, then per-page `## Heading` + YAML block (`title:` / `url:` / `description:`) + raw MDX; Verified sample=platform.claude.com/docs/llms-full.txt[^7] — https://llms-explorer.com/reference/spec/#32-llms-fulltxt-not-in-the-spec-and-three-grammars
- [parameter] Cloudflare: Page block=YAML frontmatter (`description:` / `title:` / `image:`), a "Documentation Index" blockquote pointing at the covering `//llms.txt`, `# Title`, a `[View as Markdown](…/index.md)` line, body; Verified sample=developers.cloudflare.com/llms-full.txt (57 MB)[^6] — https://llms-explorer.com/reference/spec/#32-llms-fulltxt-not-in-the-spec-and-three-grammars
- [parameter] Firecrawl generators: Page block=pages delimited by `<\|firecrawl-page-N-lllmstxt\|>`; Verified sample=create-llmstxt-py[^10] — https://llms-explorer.com/reference/spec/#32-llms-fulltxt-not-in-the-spec-and-three-grammars
- [parameter] Reference `llms_txt2ctx`: Behaviour=regex-parse, fetch every link, emit XML ``; `--optional True` includes the Optional section; removed from the proposal in v2; Evidence=[^24][^2] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] LangChain `mcpdoc` (MCP): Behaviour=`list_doc_sources` + `fetch_docs`; the *agent* decides which links to follow; allowlists only the llms.txt's domain; Evidence=[^25] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] Claude Code: Behaviour=Anthropic publishes its docs index and points the agent at it; Ahrefs' logs show the `Claude-Code` UA out-fetching every AI retrieval bot bar two (statespace-indexer, GPTBot); no documented *automatic* lookup — it is fetched when directed; Evidence=[^14][^26] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] Cursor `@Docs`: Behaviour=crawls URLs; "cannot recognise llms.txt" request acknowledged (Jun 2025), no documented support; >50–60k tokens unstable; its own llms.txt once redirected to an HTML app shell; Evidence=[^17][^27] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] Windsurf, Copilot: Behaviour=`@docs` is a curated list; Copilot feature request unanswered as of Jul 2026; Evidence=[^28][^29] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] ChatGPT, Perplexity, Google: Behaviour=no statements of use; logs ≈ 0 requests; Google: "You don't need to create new machine readable files"; Evidence=[^12][^13][^14] — https://llms-explorer.com/reference/spec/#5-how-consumers-actually-use-it
- [parameter] `robots.txt`: Job=access control; now also carries Cloudflare **Content Signals** (`Content-Signal: search=yes, ai-input=…, ai-train=no`, 2025-09-24); Do AI bots fetch it?=**full treatment: `references/robots-txt.md` (RFC 9309) and `references/robots-txt-content-signals.md`** — yes, thousands of times per site; Content Signals: Google says "no effects whatsoever"[^36][^37] — https://llms-explorer.com/reference/spec/#7-related-files
- [parameter] `sitemap.xml`: Job=exhaustive inventory; no `.md` versions, no external links; Do AI bots fetch it?=yes (ClaudeBot, GPTBot, Bingbot)[^38] — https://llms-explorer.com/reference/spec/#7-related-files
- [parameter] `llms.txt`: Job=curated navigation for agents pointed at it — **not an access-control or TDM opt-out mechanism**; see `eu-ai-act-tdm-opt-out` for why it does not count as an EU rights reservation; Do AI bots fetch it?=~0 speculative fetches; agents when directed[^14] — https://llms-explorer.com/reference/spec/#7-related-files
- [parameter] `ai.txt`: Job=opt-out preferences (IETF draft) — see `references/ai-txt.md` for the full disambiguation (5-way name collision, adoption evidence, authoring); Do AI bots fetch it?=397 instances found May 2026[^39] — https://llms-explorer.com/reference/spec/#7-related-files
- [parameter] `agents.md` / `/.well-known/ucp`: Job=Shopify's agent-commerce surface, auto-shipped to every Liquid storefront. Since 2026-05-28 `/agents.md` is **canonical** and `/llms.txt` + `/llms-full.txt` are alternate URLs that **mirror it** — so a Shopify `/llms.txt` is usually not a spec-shaped index. Full detail: `references/agents-md.md`, `references/ucp-protocol.md`; Do AI bots fetch it?=yes… — https://llms-explorer.com/reference/spec/#7-related-files
- [parameter] `robots.txt` + `License:` → RSL XML: Job=**licensing terms and price** (`ai-train`/`ai-input`/`search`, pay-per-crawl/-inference) — the layer llms.txt has no vocabulary for; see `rsl-really-simple-licensing.md`; Do AI bots fetch it?=no AI-company commitment; 3 of 130 sampled domains deploy it (2026-09-02) — https://llms-explorer.com/reference/spec/#7-related-files
- [definition] llms.txt: the spec and its grammars — Spec v2, llms-full grammars, discovery, consumers. — https://llms-explorer.com/reference/spec/#llmstxt-the-spec-and-its-grammars
- [definition] 1. What it is, in one paragraph — A markdown file — `/llms.txt` at a site root or **at any subpath** — that gives a language model a curated, priority-ordered map of a site's LLM-friendly content: an H1, a blockquote summary, optional prose, then H2 sections of `- [name](url): description` links.[^1] The links should point at… — https://llms-explorer.com/reference/spec/#1-what-it-is-in-one-paragraph
- [definition] 2. The spec, v2 — Structure, in order (verbatim from llmstxt.org):[^1] — https://llms-explorer.com/reference/spec/#2-the-spec-v2
- [definition] 3.2 llms-full.txt — not in the spec, and three grammars — `llms-full.txt` (the whole docset inlined into one markdown file) appears nowhere in the v1/v2 spec text or the repo README.[^7] Mintlify says it "was developed by Mintlify in collaboration with customer Anthropic";[^8] Lab451 dates its popularisation to early 2025.[^9] There is **no single… — https://llms-explorer.com/reference/spec/#32-llms-fulltxt-not-in-the-spec-and-three-grammars
- [definition] 3.3 `.md` twins — Mintlify, Fern, GitBook and ReadMe all serve a `.md` twin per page and link them from llms.txt "so AI tools can fetch the Markdown version of each page directly".[^4][^11] Mintlify's twins begin with a blockquote — `> ## Documentation Index` / `Fetch the complete documentation index at… — https://llms-explorer.com/reference/spec/#33-md-twins
## Generation tooling
- [parameter] Docs on Mintlify / GitBook / ReadMe / Fern: Use=nothing — it is automatic; Emits=llms.txt (+ full on Mintlify/GitBook) + `.md` twins — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] Docusaurus, MkDocs, VitePress, Starlight, Sphinx, Nuxt: Use=the framework plugin (table §3); Emits=llms.txt + llms-full.txt (+ `.md`, `llms-small.txt` on Starlight) — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] A live site you do not own: Use=crawl-based generator (§4) — `create-llmstxt-py`, `dotenvx/llmstxt`, or your own sitemap→markdown pipeline; Emits=llms.txt (+ full) with **extracted or AI-written** descriptions — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] WordPress: Use=Yoast ≥25.3 / Rank Math / AIOSEO (§5); Emits=llms.txt only (AIOSEO Pro adds full + markdown posts) — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] Webflow / Framer: Use=host a hand-written file; Emits=whatever you upload — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] Any Cloudflare-proxied HTML site: Use="Markdown for Agents" toggle (§6); Emits=on-the-fly markdown on `Accept: text/markdown`, no llms.txt — https://llms-explorer.com/reference/tooling/#1-pick-by-situation
- [parameter] **Mintlify**: Emits=llms.txt, llms-full.txt, `.md` per page, `/.well-known/` copies, `/_llms/` split indexes; Descriptions from=frontmatter `description` (truncated at 300 chars), nav order from `docs.json`; optional `markdown.instructions` agent text; Notes=index capped at 100,000 chars → recursive `/_llms/.md`; default language/version only; hidden/noindex pages excluded; hand-written… — https://llms-explorer.com/reference/tooling/#2-docs-platforms
- [parameter] **Fern**: Emits=llms.txt (root **and per-subdirectory**), `.md` per page; **no llms-full.txt**; Descriptions from=frontmatter `description`, fallback `subtitle`; adds OpenAPI/AsyncAPI links; Notes=dropped llms-full because it "exceeded most model context windows, added heavy serving overhead, saw little use"[^2] — https://llms-explorer.com/reference/tooling/#2-docs-platforms
- [parameter] **GitBook**: Emits=llms.txt (Jan 2025), llms-full.txt + `.md` per page (Jun 2025), `/sitemap.md`, `Accept: text/markdown`; Descriptions from=auto from page structure; Notes=zero-config; no curation controls documented; full export "will be more expensive"[^3][^4] — https://llms-explorer.com/reference/tooling/#2-docs-platforms
- [parameter] **ReadMe**: Emits=llms.txt (default on, all plans), `.md` per page; **no llms-full**; Descriptions from=project title + guide/API hierarchy; Notes=a custom file from the repo root disables auto-updates; hidden pages excluded[^5] — https://llms-explorer.com/reference/tooling/#2-docs-platforms
- [parameter] **GitDoc** (vendor claim): Emits=llms.txt + llms-full.txt "for the pages you mark as priority", regenerated in the build; Descriptions from=sidebar/nav; Notes=vendor blog, 2026-05-22[^6] — https://llms-explorer.com/reference/tooling/#2-docs-platforms
- [parameter] `docusaurus-plugin-llms` (rachfop): Emits=llms.txt, llms-full.txt, optional per-page `.md`, versioned + `customLLMFiles`; Input=source tree at `postBuild`; Descriptions / ordering=frontmatter → first heading → site fallback; `includeOrder` globs; Maturity & limits=144★, MIT; not run in `docusaurus start`; image rewrite only for bundled assets[^7] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `@signalwire/docusaurus-plugin-llms-txt`: Emits=llms.txt, `.md`, optional full; Input=**built HTML** (rehype/remark); Descriptions / ordering=manual `sections[].description`, `autoSectionDepth`; Maturity & limits=v1.2.2, ~10 months stale; ENOENT / "processed 0 documents" bug[^8][^9] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] Docusaurus core: Emits=none; Input=—; Descriptions / ordering=—; Maturity & limits=issue #10899 open since Feb 2025[^10] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `mkdocs-llmstxt` (pawamoy): Emits=llms.txt, `.md`, optional `full_output`; Input=built HTML → BeautifulSoup → Markdownify; Descriptions / ordering=`sections:` dict with per-file descriptions; Maturity & limits=130★, v0.5.x, **maintenance mode, seeking maintainer**; needs `site_url`; mkdocstrings `show_source` mangles tables/code in the full file[^11][^12] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `vitepress-plugin-llms` (okineadev): Emits=llms.txt, llms-full.txt, `.md`; Input=VitePress source; Descriptions / ordering=frontmatter `description`; `` / `` tags; Maturity & limits=394★; used by Vite, Vue, Vitest, Rolldown; relative URLs break under redirects/domain moves[^13] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `starlight-llms-txt` (delucis): Emits=llms.txt, llms-full.txt, **llms-small.txt**; Input=Astro Starlight; Descriptions / ordering=`projectName`, `description`, `details`, `optionalLinks`, `customSets`, `promote`/`demote`; `minify` strips asides; Maturity & limits=110★, docs updated Aug 2026; needs `site`[^14] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `sphinx-llms-txt` (jdillard): Emits=llms.txt (markdown), llms-full.txt (**reStructuredText**); Input=Sphinx build; Descriptions / ordering=toctree titles; `llms_txt_summary`, `llms_txt_exclude`, `llms_txt_full_max_size`; Maturity & limits=v0.7.1; full file is RST; points to NVIDIA `sphinx-llm`[^15] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `nuxt-llms` / Nuxt Content: Emits=llms.txt (~5K tokens), opt-in llms-full.txt (~1M+ tokens); Input=Nuxt Content, runtime hooks; Descriptions / ordering=`sections` in `nuxt.config`; Maturity & limits=first-party; full file explicitly for 200K+-context tools[^16] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] Next.js / Nextra: Emits=hand-rolled `app/llms.txt/route.ts` (force-static or dynamic); `next-llms-txt` adds per-page `.md` endpoints; Input=components; Descriptions / ordering="reads and parses readable text"; Maturity & limits=discussion #80692 unresolved; no Nextra built-in found (tentative)[^17][^18] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] `llms-txt-action` (demodrive-ai): Emits=llms.txt, llms-full.txt, `.md`; Input=built HTML dir + sitemap.xml; Descriptions / ordering=local/offline or cloud LLM summaries via LiteLLM (default GPT-4o); Maturity & limits=16★; needs `--dirty` with `mkdocs gh-deploy`[^19] — https://llms-explorer.com/reference/tooling/#3-static-site-generator-plugins
- [parameter] Firecrawl `/llmstxt` API + llmstxt.firecrawl.dev: What it does=URL → async job → llms.txt (+ full); `maxUrls` 1–100 (default 10), 1 credit/URL, public pages only, 5,000-URL alpha cap; Limits=**deprecated in favour of the main endpoints** (page carries no date; still up); users pointed to the Python repo[^20][^21] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] `create-llmstxt-py` (Firecrawl, 320★): What it does=`/map` → scrape each page to markdown (batches of 10; failures skipped, no retry) → GPT-4o-mini writes a 3–4-word title + 9–10-word description → flat llms.txt; llms-full.txt concatenates under `<\|firecrawl-page-N-lllmstxt\|>`; Limits=default 20 URLs; memory issues on large sites; **sections are not inferred**; descriptions are AI-written and… — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] `dotenvx/llmstxt` (147★, BSD-3): What it does=sitemap.xml → `- [Title](url): description` bullets; `--include-path` / `--exclude-path` globs; `--replace-title` regex; Limits=llms.txt only; titles extracted from HTML; description derivation undocumented[^23] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] Jina Reader `r.jina.ai/`: What it does=headless Chrome or curl engine → Readability → Turndown; headers `x-respond-with`, `x-target-selector`, `x-retain-links`, `x-max-tokens`, `x-markdown-chunking`; Limits=per-page cleaner, no site/llms.txt mode; anonymous traffic rate-limited[^24] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] Screaming Frog v24.3: What it does=per-page `.md` via a Readability.js + Turndown custom-JS snippet; llms.txt via n8n/CSV converters; Limits=no native llms.txt export; thin pages return nothing; JS rendering slow[^25][^26] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] `plainsignal/llmstxt` Chrome extension: What it does=llms.txt + one `.md` per page + zip from sitemap or rendered DOM; meta description as blockquote; Limits=10★, HTTPS only[^27] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] SEO-tool generators (SEOmator etc.): What it does=robots.txt → sitemap discovery, index-sitemap expansion, LLM-written title+description per URL; Limits=vendor-claimed mechanics only[^28] — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] llms-text.com generator/validator: What it does=crawls a domain and exports llms.txt + llms-full.txt ("deep-crawls up to 50 subpages"); validator checks syntax, links, UTF-8, headers; Limits=vendor; its guidance: 10–20 evergreen URLs, 4–7 H2s, 10–20-word descriptions, index under 10 KB, `Content-Type: text/plain — https://llms-explorer.com/reference/tooling/#4-crawl-based-generators-sites-you-do-not-own
- [parameter] Yoast SEO ≥25.3 (2025-06-10): Emits=llms.txt only, regenerated weekly; Descriptions=custom excerpt only — **no description otherwise**; 5 latest posts/pages/CPT (≤12 months, cornerstone first) + top-5 taxonomies; Limits=5-item cap; markdown chars escaped; a static file wins over the dynamic one[^31][^32] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] Rank Math: Emits=llms.txt only; Descriptions="intro text"; post types/taxonomies, limit default 100; custom lines; Limits=no full[^33] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] AIOSEO: Emits=llms.txt (free); llms-full.txt + markdown post conversion (Pro); Descriptions=site title/tagline; per-post-type limits, exclusions; Limits=paywall[^34] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] `website-llms-txt`, `llms-full-txt-generator`: Emits=llms.txt (+ full); Descriptions=titles + SEO-plugin descriptions; honour noindex; Limits=one shipped a broken-access-control CVE fix[^35] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] Joost de Valk "Markdown Alternate": Emits=`` + `.md` URLs per post; Descriptions=—; Limits=negotiation, not an index[^36] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] Webflow / Framer: Emits=host an uploaded file (Framer: Pro/Enterprise "Hosting → Files"); a Framer marketplace plugin scans the CMS; Descriptions=manual; Limits=no generation[^37][^38] — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [parameter] Shopify (Apr–May 2026, silent): Emits=auto `/llms.txt`, `/llms-full.txt`, `/agents.md`, `/sitemap_agentic_discovery.xml`, `/.well-known/ucp` — **Liquid storefronts only** (headless Hydrogen/Oxygen get `/.well-known/ucp` alone); Descriptions=boilerplate: H1 store name, Shop-skill pitch, UCP + MCP endpoints, `/collections/all`, policies — **no contact details** (deliberately omitted; the file is… — https://llms-explorer.com/reference/tooling/#5-cms-and-site-builders
- [definition] Generation tooling — Generators compared; why extractive descriptions win. — https://llms-explorer.com/reference/tooling/#generation-tooling
- [definition] 6. Edge content negotiation — Cloudflare "Markdown for Agents" (2026-02-12; Pro/Business/Enterprise; zone toggle under AI Crawl Control): on `Accept: text/markdown` the edge converts HTML → markdown (body + meta-derived YAML frontmatter + JSON-LD, nav/header/footer/scripts dropped) and returns `Content-Type: text/markdown… — https://llms-explorer.com/reference/tooling/#6-edge-content-negotiation
## Usage: serving, discovering and reading llms files
- [parameter] `Content-Type`: Value=`text/markdown; charset=utf-8`; Why=attribute H2; `text/plain` is tolerated, HTML is a High — https://llms-explorer.com/reference/usage/#1-serving
- [parameter] `X-Markdown-Tokens`: Value=`bytes // 4` — the same estimator `manifest.json` uses; Why=H4: cost known before fetch — https://llms-explorer.com/reference/usage/#1-serving
- [parameter] `Link`: Value=`; rel="describedby"` — the index that covers this file; Why=H3, spec v2 discovery — https://llms-explorer.com/reference/usage/#1-serving
- [parameter] `hub_docset_index(key)`: the docset's `llms.txt` (or `llms-small.txt`, `llms-facts.txt`, `manifest.json`, `/llms.txt`), with served URLs — https://llms-explorer.com/reference/usage/#5-claude-code-and-mcp
- [parameter] `hub_query_docset(key, q, mode=semantic\|keyword\|hybrid, layer=auto\|facts\|raw)`: ranked units or chunks; `layer=auto` prefers the facts layer — https://llms-explorer.com/reference/usage/#5-claude-code-and-mcp
- [parameter] `hub_llms_full_read(key, page=…)` or `(offset, limit)`: one page or a slice of a mirrored `llms-full.txt` — https://llms-explorer.com/reference/usage/#5-claude-code-and-mcp
- [parameter] `hub_llms_full_list(query, category, status, min_pages)`: which sites publish a full file, with sizes — https://llms-explorer.com/reference/usage/#5-claude-code-and-mcp
- [definition] Usage: serving, discovering and reading llms files — The headers to send, the .md twins to publish, how a reader discovers the family, how an agent reads an index, and how Claude Code and the hub MCP tools consume one. — https://llms-explorer.com/reference/usage/#usage-serving-discovering-and-reading-llms-files
- [definition] 1. Serving — Every markdown file in the family is served with: — https://llms-explorer.com/reference/usage/#1-serving
- [definition] 2. Markdown twins — Every page in the content sections — reference, examples, blog — has a clean-markdown twin at the same route with `.md` appended: `/reference/usage/` → `/reference/usage.md`. — https://llms-explorer.com/reference/usage/#2-markdown-twins
- [definition] 4. Reading an index — The v2 consumption model: *view or search the index, then follow the relevant links; the detail lives behind the links and is fetched only when needed.* As a procedure: 1. Read the H1 and blockquote — is this the product you meant? — https://llms-explorer.com/reference/usage/#4-reading-an-index
- [definition] 5. Claude Code and MCP — Claude Code fetches an llms file when directed — Anthropic publishes its own docs index and points the agent at it, and the `Claude-Code` user agent shows up in server logs ahead of every AI retrieval bot but two. The pattern is a URL in a prompt or a `CLAUDE.md`, not automatic lookup. — https://llms-explorer.com/reference/usage/#5-claude-code-and-mcp
## Worked example: the Anthropic hiring concept pack
- [definition] Worked example: the Anthropic hiring concept pack — How the Anthropic hiring concept pack in this repo was researched, structured, and source-rated — a worked example of the research-to-concept-pack workflow. — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#worked-example-the-anthropic-hiring-concept-pack
- [definition] The question — Could this repo's own research-to-concept-pack workflow produce something worth trusting on a real subject, not just a demo subject? — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#the-question
- [definition] Gathering the sources — The research step used `/dr` — firecrawl search first, with scrape as the fallback when a search snippet didn't carry enough of the page to cite confidently. — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#gathering-the-sources
- [definition] Why the pack and the facts live in separate files — `llms-anthropic-hiring.txt` is meant to be read narratively: a 10-topic index (recruiter screen, technical assessment, system design, values interview, role variations, timeline, compensation, preparation, culture signals, best practices), then a section per topic with enough prose to orient a… — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#why-the-pack-and-the-facts-live-in-separate-files
- [definition] Rating confidence — The pack-level rating stamped in `llms.txt` — "High confidence" — follows a specific rule: it holds when three or more independently produced sources agree on the same shape of the process and none contradicts it. — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#rating-confidence
- [definition] How an agent uses it — Index, then topic, then facts. `llms.txt` carries one entry describing the pack's scope, audience, and confidence rating in a few lines — cheap enough to scan on every query. — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#how-an-agent-uses-it
- [definition] What to copy for your own subject — Name the subject concretely enough to check. — https://llms-explorer.com/examples/anthropic-hiring-worked-example/#what-to-copy-for-your-own-subject
## Which layer answers which question
- [parameter] Orientation before any retrieval: what does this site cover, where do I start: layer=`llms.txt` (≤ 10 KB) then ≤ 2 hops to a `.md` twin; cost class=~3k tokens, 3 requests, 0 embeddings; recipe=recipe-01 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Orientation on a site whose index split into sections (`## Sections` present): layer=split root: root index → `/llms.txt` → page; cost class=~3–5k tokens, 3–4 requests; recipe=recipe-02 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] An exact token: an env var, a flag, a header name, an error string: layer=keyword layer (`mode="keyword"`, FTS5/BM25) over `llms-facts.txt`; cost class=microseconds, 0 model tokens, 0 embeddings; recipe=recipe-03 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] A paraphrased question, or mixed / unsure whether the words match the source: layer=hybrid (`mode="hybrid"`, RRF over keyword + vector), or vector alone (`layer="facts"`); cost class=1 embedding, 0 generation tokens; recipe=recipe-04 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] An agent that must find the right page from an MCP client without a search index: layer=index-first via `hub_docset_index` → `sections` → section index → page; cost class=~2k tokens read per hop, 0 embeddings; recipe=recipe-05 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] A scripted check or query from a shell or a CI step: layer=the `llmsx` CLI (today: the hub scripts it wraps); cost class=seconds; 0 model tokens for lint / keyword; recipe=recipe-06 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Citation-grade answers inside your own RAG store: layer=`llms-facts.txt` units, one document each, `url#anchor` as metadata; cost class=1 embedding per unit at ingest; ~845k tokens for a 191-page site; recipe=recipe-07 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Keeping a published file honest on every push: layer=the lint as a GitHub Action gate (exit 1 on High); cost class=~10 s per file; network only with `--check-links`; recipe=recipe-08 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Serving the files so agents and the lint can find them: layer=headers: `text/markdown`, `X-Markdown-Tokens`, `Link: rel="describedby"`, `rel="alternate"` on HTML; cost class=one config block; verify with `curl -I`; recipe=recipe-09 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Whole-corpus reasoning, offline and private, within a token budget: layer=a local hub: Ollama + indexer + keyword layer + `llms_serve.py`; `llms-small.txt` for budgeted reads; cost class=one machine; ~50k tokens per small read, 0 API spend; recipe=recipe-10 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] One concept across many sources, disagreements visible: layer=a topical file (`/t//`) built from a fact pool; cost class=minutes to build; `--no-embed` for 0 embeddings; recipe=recipe-11 — https://llms-explorer.com/examples/decision-table/#the-table
- [parameter] Disambiguation: which sense of a word this family means, and its aliases: layer=`llms-vocabulary.txt` senses and `aka:` expansion before FTS5; cost class=free: string match, 0 model tokens; recipe=recipe-12 — https://llms-explorer.com/examples/decision-table/#the-table
- [definition] Which layer answers which question — The decision table for the cookbook: match the shape of your question to the cheapest llms layer that answers it, then open the recipe. — https://llms-explorer.com/examples/decision-table/#which-layer-answers-which-question
- [definition] When the table is the wrong tool — If the question is "is this file any good", none of these rows apply — that is the [lint](/reference/passes/), not a retrieval. — https://llms-explorer.com/examples/decision-table/#when-the-table-is-the-wrong-tool
## Recipe 01 — Two hops with requests
- [snippet] Steps: import re, requests — `import re, requests` — https://llms-explorer.com/examples/recipe-01/#steps
- [snippet] Expected output: https://llms-explorer.pages.dev/examples/recipe-09/ https://llms-explorer.pages.dev/examples/recipe-09.md — `https://llms-explorer.pages.dev/examples/recipe-09/ https://llms-explorer.pages.dev/examples/recipe-` — https://llms-explorer.com/examples/recipe-01/#expected-output
- [definition] Recipe 01 — Two hops with requests — Read a site's llms.txt, pick a page by its description, fetch the .md twin, answer. The baseline every other recipe is measured against. — https://llms-explorer.com/examples/recipe-01/#recipe-01-two-hops-with-requests
## Recipe 02 — Split root: follow a section index
- [snippet] Steps: import re, requests — `import re, requests` — https://llms-explorer.com/examples/recipe-02/#steps
- [snippet] Expected output: ('Agent Sdk', 'agent-sdk/llms.txt') — `('Agent Sdk', 'agent-sdk/llms.txt')` — https://llms-explorer.com/examples/recipe-02/#expected-output
- [definition] Recipe 02 — Split root: follow a section index — When the root llms.txt has a ## Sections block, let the counts on each section line decide which section index to fetch before touching a page. — https://llms-explorer.com/examples/recipe-02/#recipe-02-split-root-follow-a-section-index
## Recipe 03 — Keyword layer from Claude Code
- [snippet] Steps: hub_query_docset(docset="codeclaudecom__codeclaudecom", question="CLAUDE_CODE_SYNC_SKILLS", mode="keyword", top=3) — `hub_query_docset(docset="codeclaudecom__codeclaudecom", question="CLAUDE_CODE_SYNC_SKILLS", mode="ke` — https://llms-explorer.com/examples/recipe-03/#steps
- [snippet] Steps: { — `{` — https://llms-explorer.com/examples/recipe-03/#steps
- [snippet] Steps: hub_llms_full_read(key="code.claude.com__docs", page="https://code.claude.com/docs/en/env-vars") — `hub_llms_full_read(key="code.claude.com__docs", page="https://code.claude.com/docs/en/env-vars")` — https://llms-explorer.com/examples/recipe-03/#steps
- [snippet] Steps: { — `{` — https://llms-explorer.com/examples/recipe-03/#steps
- [snippet] Steps: .venv/bin/python scripts/docset_indexer.py keyword codeclaudecom__codeclaudecom "CLAUDE_CODE_SYNC_SKILLS" --layer facts — `.venv/bin/python scripts/docset_indexer.py keyword codeclaudecom__codeclaudecom "CLAUDE_CODE_SYNC_SK` — https://llms-explorer.com/examples/recipe-03/#steps
- [definition] Recipe 03 — Keyword layer from Claude Code — Find an exact token — an env var, a flag, an error string — with hub_query_docset(mode=\"keyword\"), then open the page it came from. Zero model tokens. — https://llms-explorer.com/examples/recipe-03/#recipe-03-keyword-layer-from-claude-code
## Recipe 04 — Hybrid: keyword and vector fused
- [snippet] Steps: hub_query_docset( — `hub_query_docset(` — https://llms-explorer.com/examples/recipe-04/#steps
- [snippet] Steps: { — `{` — https://llms-explorer.com/examples/recipe-04/#steps
- [snippet] Steps: .venv/bin/python scripts/docset_indexer.py query codeclaudecom__codeclaudecom "which environment variable downloads my — `.venv/bin/python scripts/docset_indexer.py query codeclaudecom__codeclaudecom "which environment v` — https://llms-explorer.com/examples/recipe-04/#steps
- [definition] Recipe 04 — Hybrid: keyword and vector fused — For a paraphrased or uncertain question, mode=\"hybrid\" runs the keyword and vector legs and fuses them with reciprocal-rank fusion; legs == 2 tells you both agreed on a hit. — https://llms-explorer.com/examples/recipe-04/#recipe-04-hybrid-keyword-and-vector-fused
## Recipe 05 — Index-first agent over MCP
- [snippet] Steps: hub_docset_index(docset="codeclaudecom__codeclaudecom") — `hub_docset_index(docset="codeclaudecom__codeclaudecom")` — https://llms-explorer.com/examples/recipe-05/#steps
- [snippet] Steps: { — `{` — https://llms-explorer.com/examples/recipe-05/#steps
- [snippet] Steps: hub_docset_index(docset="codeclaudecom__codeclaudecom", file="agent-sdk/llms.txt") — `hub_docset_index(docset="codeclaudecom__codeclaudecom", file="agent-sdk/llms.txt")` — https://llms-explorer.com/examples/recipe-05/#steps
- [snippet] Steps: { — `{` — https://llms-explorer.com/examples/recipe-05/#steps
- [definition] Recipe 05 — Index-first agent over MCP — hub_docset_index → read sections → the section's llms.txt → the page. The pattern a concept-tree node page uses to find a source without any search index. — https://llms-explorer.com/examples/recipe-05/#recipe-05-index-first-agent-over-mcp
## Recipe 06 — The llmsx CLI
- [snippet] Steps: llmsx lint ./docs/llms.txt --json — `llmsx lint ./docs/llms.txt --json` — https://llms-explorer.com/examples/recipe-06/#steps
- [snippet] Steps: llmsx query code.claude.com "CLAUDE_CODE_SYNC_SKILLS" --mode keyword — `llmsx query code.claude.com "CLAUDE_CODE_SYNC_SKILLS" --mode keyword` — https://llms-explorer.com/examples/recipe-06/#steps
- [snippet] Steps: llmsx export mirrors/code.claude.com.md — `llmsx export mirrors/code.claude.com.md` — https://llms-explorer.com/examples/recipe-06/#steps
- [snippet] Steps: llmsx tree show "llms.txt" — `llmsx tree show "llms.txt"` — https://llms-explorer.com/examples/recipe-06/#steps
- [snippet] Expected output: $ llmsx lint site/dist/llms.txt site/dist/llms-facts.txt --json | jq -c '.[] | {file, high: .counts.high}' — `$ llmsx lint site/dist/llms.txt site/dist/llms-facts.txt --json | jq -c '.[] | {file, high: .counts.` — https://llms-explorer.com/examples/recipe-06/#expected-output
- [definition] Recipe 06 — The llmsx CLI — Lint, query, export and inspect the tree from a shell: the llmsx commands and the hub scripts each one wraps today. — https://llms-explorer.com/examples/recipe-06/#recipe-06-the-llmsx-cli
## Recipe 07 — Facts into a RAG store
- [snippet] Steps: import hashlib, re — `import hashlib, re` — https://llms-explorer.com/examples/recipe-07/#steps
- [snippet] Steps: import chromadb, requests — `import chromadb, requests` — https://llms-explorer.com/examples/recipe-07/#steps
- [snippet] Expected output: {'type': 'parameter', 'url': 'https://code.claude.com/docs/en/admin-setup', 'anchor': 'set-up-claude-code-for-your-organ — `{'type': 'parameter', 'url': 'https://code.claude.com/docs/en/admin-setup', 'anchor': 'set-up-claude` — https://llms-explorer.com/examples/recipe-07/#expected-output
- [definition] Recipe 07 — Facts into a RAG store — Parse llms-facts.txt with UNIT_RE, one document per unit with its url#anchor as metadata, embed with mxbai-embed-large — and never mix it with a 768-dimension model. — https://llms-explorer.com/examples/recipe-07/#recipe-07-facts-into-a-rag-store
## Recipe 08 — GitHub Action lint gate
- [snippet] Steps: name: llms lint — `# .github/workflows/llms-lint.yml` — https://llms-explorer.com/examples/recipe-08/#steps
- [snippet] Expected output: [ — `[` — https://llms-explorer.com/examples/recipe-08/#expected-output
- [snippet] Expected output: {"pass": "P0", "attr": "I6", "severity": "high", "line": 0, — `{"pass": "P0", "attr": "I6", "severity": "high", "line": 0,` — https://llms-explorer.com/examples/recipe-08/#expected-output
- [definition] Recipe 08 — GitHub Action lint gate — Fail a pull request on any High finding in your llms files, and annotate the offending lines from the lint's JSON. — https://llms-explorer.com/examples/recipe-08/#recipe-08-github-action-lint-gate
## Recipe 09 — Serving with the right headers
- [snippet] Steps: types { text/markdown md; } — `# inside the server {} block` — https://llms-explorer.com/examples/recipe-09/#steps
- [snippet] Steps: /*.md — `/*.md` — https://llms-explorer.com/examples/recipe-09/#steps
- [snippet] Steps: curl -sI https://docs.example.com/reference/attributes.md | grep -iE '^(content-type|x-markdown-tokens|link):' — `curl -sI https://docs.example.com/reference/attributes.md | grep -iE '^(content-type|x-markdown-toke` — https://llms-explorer.com/examples/recipe-09/#steps
- [snippet] Expected output: content-type: text/markdown; charset=utf-8 — `content-type: text/markdown; charset=utf-8` — https://llms-explorer.com/examples/recipe-09/#expected-output
- [snippet] Expected output: link: ; rel="alternate"; type="text/markdown" — `link: ; rel="alternate"; type="text/markdown"` — https://llms-explorer.com/examples/recipe-09/#expected-output
- [definition] Recipe 09 — Serving with the right headers — nginx and Cloudflare _headers blocks that serve .md twins as text/markdown with X-Markdown-Tokens and the two Link relations, verified with curl -I. — https://llms-explorer.com/examples/recipe-09/#recipe-09-serving-with-the-right-headers
## Recipe 10 — A local hub in miniature
- [snippet] Steps: ollama pull mxbai-embed-large — `ollama pull mxbai-embed-large` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Steps: python3 scripts/llms_acquire.py https://code.claude.com/docs mirrors/code.claude.com.md — `python3 scripts/llms_acquire.py https://code.claude.com/docs mirrors/code.claude.com.md` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Steps: PYTHONPATH=scripts .venv/bin/python -m docset_refine all --no-units mirrors/code.claude.com.md — `PYTHONPATH=scripts .venv/bin/python -m docset_refine all --no-units mirrors/code.claude.com.md` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Steps: .venv/bin/python scripts/docset_indexer.py index mirrors/code.claude.com.md --name code.claude.com — `.venv/bin/python scripts/docset_indexer.py index mirrors/code.claude.com.md --name code.claude.com` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Steps: .venv/bin/python scripts/llms_serve.py --host 127.0.0.1 --port 8788 — `.venv/bin/python scripts/llms_serve.py --host 127.0.0.1 --port 8788` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Steps: .venv/bin/python scripts/docset_indexer.py keyword code.claude.com "CLAUDE_CODE_SYNC_SKILLS" --layer facts --mode phrase — `.venv/bin/python scripts/docset_indexer.py keyword code.claude.com "CLAUDE_CODE_SYNC_SKILLS" --layer` — https://llms-explorer.com/examples/recipe-10/#steps
- [snippet] Expected output: HTTP/1.0 200 OK — `HTTP/1.0 200 OK` — https://llms-explorer.com/examples/recipe-10/#expected-output
- [definition] Recipe 10 — A local hub in miniature — Ollama, the docset indexer, a keyword layer and llms_serve.py on one machine: the whole retrieval stack for one family, offline and private. — https://llms-explorer.com/examples/recipe-10/#recipe-10-a-local-hub-in-miniature
## Recipe 11 — Building a topical file
- [snippet] Steps: PYTHONPATH=scripts .venv/bin/python -m docset_refine topical \ — `PYTHONPATH=scripts .venv/bin/python -m docset_refine topical \` — https://llms-explorer.com/examples/recipe-11/#steps
- [snippet] Steps: .venv/bin/python scripts/llms_lint.py check llms-topical/prompt-caching.llms/ --json — `.venv/bin/python scripts/llms_lint.py check llms-topical/prompt-caching.llms/ --json` — https://llms-explorer.com/examples/recipe-11/#steps
- [definition] Recipe 11 — Building a topical file — docset_refine topical turns a fact pool into a concept-axis llms.txt + llms-facts.txt with the subject's child concepts as sections; then /ldo --agent-test checks an agent can actually use it. — https://llms-explorer.com/examples/recipe-11/#recipe-11-building-a-topical-file
## Recipe 12 — Reading a vocabulary
- [snippet] Steps: import re, requests — `import re, requests` — https://llms-explorer.com/examples/recipe-12/#steps
- [snippet] Expected output: ("/llms.txt" OR "/llmstxt" OR "llms.txt") discovery — `("/llms.txt" OR "/llmstxt" OR "llms.txt") discovery` — https://llms-explorer.com/examples/recipe-12/#expected-output
- [snippet] Expected output: - **cookie** [web.cookie] · [folklore.cookie-monster] · [food.cookie]: … — `- **cookie** [web.cookie] · [folklore.cookie-monster] · [food.cookie]: …` — https://llms-explorer.com/examples/recipe-12/#expected-output
- [definition] Recipe 12 — Reading a vocabulary — Expand a query through a family's aka: list before the FTS5 lookup, and pin the sense the family means. Free: string matching, no model. — https://llms-explorer.com/examples/recipe-12/#recipe-12-reading-a-vocabulary
## A Closed-Loop System for Autonomous Skill-Knowledge Acquisition
- [snippet] 2. System architecture: ┌──────────────────────────────────────────────────────────┐ — ` ┌──────────────────────────────────────────────────────────┐` — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#2-system-architecture
- [snippet] 3.1 Node schema: { — `{` — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#31-node-schema
- [parameter] `concept-family-explorer`: Job=decide *what to build*; Scope per invocation=a whole family; Reads=the tree (coverage); Writes=nodes + parent/child links — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#2-system-architecture
- [parameter] `/dr`: Job=*build one thing* well; Scope per invocation=a single concept; Reads=the web; Writes=one node + one skill — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#2-system-architecture
- [parameter] concept tree: Job=*remember everything*; Scope per invocation=the whole hub; Reads=—; Writes=— (it is the store) — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#2-system-architecture
- [parameter] Node.js Concurrency Internals: researchedAt=2026-06-01; sources=11 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] JS/TS Runtimes (Deno, Bun, Edge) & WinterTC: researchedAt=2026-06-01; sources=17 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Native TypeScript, Permission Model & SEA: researchedAt=2026-06-01; sources=9 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Module Resolution & ESM/CJS Interop: researchedAt=2026-06-02; sources=7 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js HTTP & Networking: researchedAt=2026-06-02; sources=10 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Modern Batteries-Included Built-ins: researchedAt=2026-06-02; sources=19 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Async Control-Flow, Errors & Context: researchedAt=2026-06-02; sources=8 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Built-in Test Runner (node:test): researchedAt=2026-06-02; sources=6 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Package Management & Supply-Chain: researchedAt=2026-06-02; sources=13 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] Node.js Production Diagnostics & Profiling: researchedAt=2026-06-02; sources=7 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [parameter] 15:09:51: Concept / skill=`mongodb-backup-restore`; child concepts=9; sources=20 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:11:05: Concept / skill=`mongodb-upgrade-paths`; child concepts=**33**; sources=20 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:12:34: Concept / skill=`mongodb-aggregation-pipeline`; child concepts=10; sources=11 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:16:04: Concept / skill=`mongodb-migration-patterns`; child concepts=10; sources=10 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:22:05: Concept / skill=`mongodb-cost-optimization`; child concepts=12; sources=— — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:23:05: Concept / skill=`mongodb-atlas-multicloud`; child concepts=12; sources=10 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:25:53: Concept / skill=`mongodb-indexes-deep`; child concepts=**17**; sources=19 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:30:15: Concept / skill=`mongodb-monitoring-observability`; child concepts=—; sources=— — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:31:25: Concept / skill=`mongodb-geospatial`; child concepts=10; sources=6 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:31:37: Concept / skill=`mongodb-atlas-stream-processing`; child concepts=12; sources=4 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:31:42: Concept / skill=`mongodb-realm-mobile-sync`; child concepts=11; sources=— — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:32:08: Concept / skill=`mongodb-atlas-triggers-functions`; child concepts=12; sources=7 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:33:10: Concept / skill=`mongodb-compliance`; child concepts=12; sources=1 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:34:27: Concept / skill=`mongodb-transactions`; child concepts=10; sources=6 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:34:29: Concept / skill=`mongodb-atlas-charts`; child concepts=11; sources=1 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 15:41:33: Concept / skill=`mongodb-capacity-planning`; child concepts=12; sources=6 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] 16:41:35: Concept / skill=`mongodb-kafka-connector`; child concepts=15; sources=18 — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [parameter] Total concept-tree nodes: Value=431; Source=`tam_concept_tree_list` (`total`) — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Total registry skills: Value=477; Source=`tam_recommend_skills` (`total`) — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] `TypeScript Expert` child concepts / sources / refresh: Value=13 / 105 / 2026-06-04; Source=`tam_concept_tree_list` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] TypeScript run sequence (1→10 spokes, 4→13, two runs): Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] `JavaScript and Node.js` sources: Value=131; Source=`tam_concept_tree_list` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Node.js new spokes + timestamps + per-node sources: Value=per §6.2 table; Source=`tam_concept_tree_search` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Node.js 10→20 spokes, 2026-06-02 saturation: Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] MongoDB cascade concepts, timestamps, child/source counts: Value=per §6.3 table; Source=`tam_concept_tree_list` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] `mongodb-upgrade-paths` 33 children / `mongodb-indexes-deep` 17: Value=33 / 17; Source=`tam_concept_tree_list` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Markdown family: 8 children, all `skillId: document-formats`, 2026-06-03: Value=8; Source=`tam_concept_tree_search` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] LLM concepts with `parentConcept` llm-models but `skillId` ai-agent-engineering: Value=—; Source=`tam_concept_tree_search` — live — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Hub-index reduction 413→187: Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Coverage-saturation threshold ≈3.2: Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Parallel-author / serial-write RMW lesson: Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Auto-builder vs manual-run collision & unpinned-build pruning: Value=—; Source=run-history — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
- [parameter] Breadth loop: Skill / tool id=`concept-family-explorer`; One-line role=Map a family, score gaps, loop `/dr` to saturation — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Depth worker: Skill / tool id=`/dr` (deep-research-build); One-line role=Research → author skill → install → cross-link → write back — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Research engine: Skill / tool id=`deep-research`; One-line role=Multi-source cited research (firecrawl/exa, web fallback) — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Shared ledger: Skill / tool id=concept tree via `tam_concept_tree_{list,get,search,upsert,link,delete}`; One-line role=Persistent graph of researched concepts — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Tree↔skill join: Skill / tool id=`skillId` field; One-line role=Many concept nodes → one hub skill — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Catalog shape: Skill / tool id=`skill-tree-architect`; One-line role=Fold saturated families into hubs; keep the scan surface flat — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [parameter] Terminal quality gate: Skill / tool id=`skill-optimizer` + `prompt-deep-optimizer`; One-line role=Audit/verify/tighten new skills after saturation — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-b-component-reference
- [definition] A Closed-Loop System for Autonomous Skill-Knowledge Acquisition — A June 2026 report traces concept-family-explorer, the concept tree, and /dr through recorded TypeScript, Node.js, and MongoDB skill builds. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#a-closed-loop-system-for-autonomous-skill-knowledge-acquisition
- [definition] Abstract (BLUF) — An LLM agent that researches one topic at a time gets *deeper*; it does not get *broader*. Depth alone leaves blind spots — the agent never learns what it does not know to ask. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#abstract-bluf
- [definition] 1. The problem: breadth is not the same as depth — Consider the ordinary way an agent acquires expertise. A user asks about TypeScript declaration files; the agent researches declaration files; the agent is now better at declaration files. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#1-the-problem-breadth-is-not-the-same-as-depth
- [definition] 2. System architecture — The three components form a loop. CFE runs the outer cycle (breadth), `/dr` runs the inner unit of work (depth), and the concept tree is the state both read from and write to. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#2-system-architecture
- [definition] 3. The concept tree: the shared-state ledger — The concept tree is the component that makes the other two work, and it is the easiest to overlook because it does no "thinking." It is a persistent graph stored in the mdb-context-hub and reached through six MCP tools: `tam_concept_tree_list`, `_get`, `_search`, `_upsert`, `_link`, and `_delete`. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#3-the-concept-tree-the-shared-state-ledger
- [definition] 3.1 Node schema — Every node is one researched concept: — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#31-node-schema
- [definition] 3.2 The tree is not the skill registry — It is tempting to assume one concept \= one skill. It is not. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#32-the-tree-is-not-the-skill-registry
- [definition] 3.3 What the tree is *for* — The ledger plays two roles that the breadth loop depends on completely: 1. **The saturation oracle.** "Is this family done?" is answerable only against a record of what is already covered. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#33-what-the-tree-is-for
- [definition] 4. `/dr`: the depth worker — `/dr` is the unit of production. Its contract is narrow on purpose: take one named topic, return one durable skill, and leave the tree richer than it found it. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#4-dr-the-depth-worker
- [definition] 5. `concept-family-explorer`: the breadth loop — CFE sits one level *above* `/dr`. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#5-concept-family-explorer-the-breadth-loop
- [definition] 5.1 The algorithm — **(1) Map the family.** Given a subject, CFE enumerates five neighborhoods around it: - the **parent** domain it belongs to, - its **siblings** (peers under the same parent), - its **sub-concepts** (children — the decomposition), - **adjacent / cross-over** fields (neighbors in other families that… — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#51-the-algorithm
- [definition] 5.2 Why the separation matters — Because CFE never does research itself, it can reason about a whole family cheaply — a tree diff plus a scoring pass — and spend the expensive web-research budget only where the score justifies it. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#52-why-the-separation-matters
- [definition] 6. Worked examples (from the live tree) — The four expansions below are real. Counts and timestamps were read from the tree while drafting this paper; see Appendix A. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#6-worked-examples-from-the-live-tree
- [definition] 6.1 The TypeScript language subtree — saturating over two runs — At the 2026-06-17 snapshot, `TypeScript Expert` (`skillId: typescript-expert`) held **13 child concepts** and carries 105 sources, last refreshed **2026-06-04**. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#61-the-typescript-language-subtree-saturating-over-two-runs
- [definition] 6.2 The Node.js core family — 10 → 20 spokes — `JavaScript and Node.js` (`skillId: javascript-nodejs`, 131 sources) anchors a family that CFE drove to core-family saturation on **2026-06-02**, *adding ten new spokes and taking the family from 10 to 20 references* *(run-history)*. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#62-the-nodejs-core-family-10-20-spokes
- [definition] 6.3 The MongoDB build cascade — a saturation run you can timestamp — The single most legible example is the MongoDB expansion of **2026-05-28**, because the `researchedAt` stamps show when the ledger recorded each skill. They do not establish worker start times, total build duration, or independent skill quality. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#63-the-mongodb-build-cascade-a-saturation-run-you-can-timestamp
- [definition] 6.4 Saturate, then fold: markdown → `document-formats` — §3.2 introduced the markdown family as proof that concepts ≠ skills. As a CFE outcome it shows the *terminal* move of a saturation run. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#64-saturate-then-fold-markdown-document-formats
- [definition] 6.5 Cross-hub placement: the LLM model layer — A last nuance the tree exposes: a concept's *conceptual* parent and its *owning* skill can differ. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#65-cross-hub-placement-the-llm-model-layer
- [definition] 7. Results — and what "meaningful" means here — The headline figures, read live while drafting: **431 concept nodes** in the tree and **477 skills** in the registry. But raw counts are the least interesting measure. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#7-results-and-what-meaningful-means-here
- [definition] 8. Limitations and failure modes — This system is not magic, and its sharp edges are worth stating plainly. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#8-limitations-and-failure-modes
- [definition] 8.1 Concurrency: the scheduled builder vs. the manual run — The most consequential failure mode: an **active scheduled CFE/auto-hub run and a manual run on the same subject collide** *(run-history)*. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#81-concurrency-the-scheduled-builder-vs-the-manual-run
- [definition] 8.2 The write-back is an unlocked read-modify-write — As noted in §4, the manifest edit during `/dr` write-back is an RMW with no lock. Parallelizing the *write* (as opposed to the research) corrupts it. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#82-the-write-back-is-an-unlocked-read-modify-write
- [definition] 8.3 Saturation is a judgment call — The stop condition is a score threshold, and the threshold (≈3.2 in one recorded run) *(run-history)* is tuned, not derived. Set it too high and families are abandoned with real gaps; too low and the loop chases diminishing concepts. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#83-saturation-is-a-judgment-call
- [definition] 8.4 Source-count is a proxy, not a guarantee — `sourcesCount` measures how many citations `/dr` gathered, not whether the synthesis is correct. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#84-source-count-is-a-proxy-not-a-guarantee
- [definition] 8.5 Provenance honesty — Several figures in this paper (the 1→10 and 10→20 spoke counts, the 413→187 index reduction, the 3.2 threshold, the parallel-author/serial-write lesson) come from **run-history, not a live tree query**, and are flagged as such. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#85-provenance-honesty
- [definition] 9. Conclusion — The design idea worth taking away is the **three-way separation of breadth, depth, and memory**, with the concept tree serving as the coverage ledger while installed skills and manifests remain durable too. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#9-conclusion
- [definition] Appendix A — Provenance of cited numbers — All figures labeled live were read via the mdb-context-hub `tam_concept_tree_*` MCP tools while drafting (2026-06-17). These records describe pipeline activity and stored metadata; they are not an independent performance or skill-quality benchmark. — https://llms-explorer.com/blog/a-closed-loop-system-for-autonomous-skill-knowledge-acquisition/#appendix-a-provenance-of-cited-numbers
## Abstracting one concept out of many docsets
- [snippet] Commands: S=~/.claude/skills/llms-concept-abstractor/scripts/concept_abstract.py — `# cwd: ~/.global-ai-hub (the skill's script; /lca wraps these steps for an agent)` — https://llms-explorer.com/blog/abstracting-one-concept/#commands
- [parameter] eval-1 with skill: Model tokens=336,335; Wall time=718 s; Grade=7/7 — https://llms-explorer.com/blog/abstracting-one-concept/#outputs
- [parameter] eval-1 baseline (ordinary tools): Model tokens=376,182; Wall time=571 s; Grade=3/7 — https://llms-explorer.com/blog/abstracting-one-concept/#outputs
- [parameter] eval-2 with skill (scope discovery): Model tokens=496,462; Wall time=2,281 s; Grade=6/6 — https://llms-explorer.com/blog/abstracting-one-concept/#outputs
- [definition] Abstracting one concept out of many docsets — Two August 2026 /lca evaluations build indexing and prompt-caching packs with lexicons and local embeddings, reporting task grades, runtime tradeoffs, and later cleanup. — https://llms-explorer.com/blog/abstracting-one-concept/#abstracting-one-concept-out-of-many-docsets
## Anchors that point nowhere
- [snippet] Commands: .venv/bin/python scripts/llms_lint.py check text-mirror/code.claude.com.llms/ \ — `# cwd: ~/.global-ai-hub` — https://llms-explorer.com/blog/anchors-that-point-nowhere/#commands
- [definition] Anchors that point nowhere — How raw-mirror heading checks corrected recorded anchor failures, and why those checks do not prove that every link works on the live rendered site. — https://llms-explorer.com/blog/anchors-that-point-nowhere/#anchors-that-point-nowhere
## Automated Auditing and Security-First Software Design
- [parameter] **SCA** (Software Composition Analysis): Question it answers=Do my dependencies have known CVEs?; Typical stage=PR + scheduled; Representative tools=Dependabot, `npm audit`, OWASP Dependency-Check, Snyk — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **SAST** (Static Application Security Testing): Question it answers=Does my source contain a vulnerable pattern?; Typical stage=PR; Representative tools=CodeQL, Semgrep, SonarQube — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **Secret scanning**: Question it answers=Did someone commit a credential?; Typical stage=Pre-commit + PR + history; Representative tools=gitleaks, trufflehog, GitHub secret scanning — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **IaC scanning**: Question it answers=Is my Terraform/K8s config insecure?; Typical stage=PR; Representative tools=Checkov, tfsec, Trivy — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **Container/image scanning**: Question it answers=Does my base image carry known CVEs?; Typical stage=Build; Representative tools=Trivy, Grype, Clair — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **DAST** (Dynamic): Question it answers=Is the *running* app exploitable?; Typical stage=Staging; Representative tools=OWASP ZAP, Burp — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **Policy as code**: Question it answers=Does this change violate an org rule?; Typical stage=PR + admission; Representative tools=OPA/Conftest, custom checks — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [parameter] **False positives**: Why it is real=SAST especially; erodes trust and trains bypass behavior; Mitigation=Advisory-first; curate rules; promote to blocking only when trusted — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#7-costs-and-honest-limits
- [parameter] **Gate friction**: Why it is real=Every blocking check is latency on the critical path of shipping; Mitigation=Fast checks on PR, slow checks scheduled; parallelize; cache — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#7-costs-and-honest-limits
- [parameter] **Tool sprawl**: Why it is real=Each scanner is config, findings, and a dashboard to maintain; Mitigation=Consolidate where possible; not every category needs a tool on day one — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#7-costs-and-honest-limits
- [parameter] **Automation's ceiling**: Why it is real=Scanners find *known* patterns and *known* CVEs — coverage of novel design flaws and business-logic abuse is incomplete; Mitigation=Keep human threat modeling and periodic pentests; automation augments, never replaces — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#7-costs-and-honest-limits
- [parameter] **Provenance is young**: Why it is real=SBOM/SLSA tooling and norms are still settling in 2026; Mitigation=Adopt incrementally; treat as a maturing capability, not a finished one — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#7-costs-and-honest-limits
- [definition] Automated Auditing and Security-First Software Design — Security design and CI auditing through a June mdb-context-hub case study, distinguishing documented controls from enforcement and coverage. — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#automated-auditing-and-security-first-software-design
- [definition] 1. The problem: security bolted on is security that fails — Security was historically the last gate before release: a penetration test in the final sprint, a manual audit against a standard, a sign-off from a team that did not write the code. Three forces have broken that model. — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#1-the-problem-security-bolted-on-is-security-that-fails
- [definition] 2. Why the usual approaches fall short — Before proposing a model, an honest review of the approaches teams actually use, and where each runs out. — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#2-why-the-usual-approaches-fall-short
- [definition] 3. Security-first design: the architectural half — "Secure by design" is older than DevSecOps. — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#3-security-first-design-the-architectural-half
- [definition] 4. Automated auditing: the pipeline half — Automated auditing moves the checklist into CI/CD and runs it on every change. The toolchain has well-defined categories, each answering a different question and fitting a different stage: — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#4-automated-auditing-the-pipeline-half
- [definition] 6. Implementation: a staged adoption path — Security-first practice fails most often from being adopted all at once. A workable sequence: 1. — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#6-implementation-a-staged-adoption-path
- [definition] 8. Conclusion and next steps — Security stops being a periodic event and becomes a property of the system when two things are true at once: the architecture is designed to be secure (least privilege, secure defaults, a known trust boundary, threat-modeled at design time), and the pipeline audits every change automatically… — https://llms-explorer.com/blog/automated-auditing-and-security-first-software-design/#8-conclusion-and-next-steps
## Building Codebases for Machine Collaborators
- [parameter] Automated documentation: What it buys the agent=Docs that cannot silently lie; mdb-tam implementation=`scripts/generate-ops-registry-doc.mjs` → `docs/operations-registry.json` (18 ops, from `server/src/lib/operations-registry.js`), `--check` CI drift gate; `generate_llm_repo_index.py`; `rotate-workflow-logs.mjs`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Documentation as architecture: What it buys the agent=A control surface the agent acts through; mdb-tam implementation=`CLAUDE.md` (authoritative, loaded every session), `AGENTS.md` (13-agent catalog over `.claude/agents/`), `GEMINI.md` (defers to `CLAUDE.md`), per-dir `README.md`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Retrieval index (curated): What it buys the agent=Judgment about which files matter and how to run them; mdb-tam implementation=`docs/high_signal_file_index.json` — 491 entries w/ `how_to_run`, `inputs/outputs`, `risk_notes`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Retrieval index (generated): What it buys the agent=Complete, regenerable coverage map; mdb-tam implementation=`docs/llm-repo-index.json` (758 files / 96 dirs) + `llm-repo-index.md`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Index integrity gate: What it buys the agent=An index that cannot rot into a trap; mdb-tam implementation=`scripts/check-doc-indexes.mjs` — validates 491 + 758 paths, CI-fails on dead paths, `--prune`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Structured server logging: What it buys the agent=Observability of runtime the agent can't watch; mdb-tam implementation=`server/src/telemetry/logger.js` — `pino` + scoped `logger.child({ scope })`; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Structured client logging: What it buys the agent=Bounded, schema'd, multi-sink error record; mdb-tam implementation=`src/background/error-log.js` (v1 schema, 1000-entry ring buffer, Sentry-lite + file sinks); Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Log-triggered remediation: What it buys the agent=A log that closes the loop; mdb-tam implementation=`error-log.js` Copilot-CLI auto-remediation path (internal repo, 180 s); Status=Active (opt-in) — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Test suite: What it buys the agent=Converts agent output from plausible to verified; mdb-tam implementation=707 tests / 4 suites (294 / 346 / 30 / 37), 3 CI workflows; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Substrate-aware harness: What it buys the agent=Extension/server logic testable headlessly; mdb-tam implementation=`test/setup.js` chrome shims; `mongodb-memory-server` graceful-skip; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] Dual-mode capability (CLI+API+MCP): What it buys the agent=One tested core, reachable by test and agent alike; mdb-tam implementation=CallCard `cli/call.js` + `routes/call.js` + `mcp/call-mcp-server.js`; corpus/reports API + thin `mcp-server/client.js`; Live Hub Toolkit `src/core` + `cli/` + native bridge; Status=Active — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [parameter] `scripts/generate-ops-registry-doc.mjs`: Practice=Automated docs; Role=Sole writer of `docs/operations-registry.json`; `--check` CI drift gate — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `server/src/lib/operations-registry.js`: Practice=Automated docs; Role=Source of truth (18 operations) for the generated doc — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `scripts/generate_llm_repo_index.py`: Practice=Automated docs; Role=Regenerates the machine-readable repo index — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `scripts/rotate-workflow-logs.mjs`: Practice=Automated docs; Role=Bounds append-only journals (keeps the latest version sections in the root; archives older sections to `docs/archive/`) — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `CLAUDE.md` (root + `~/.claude/CLAUDE.md`): Practice=Docs as architecture; Role=Authoritative rules, loaded in full every session — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `AGENTS.md` + `.claude/agents/`: Practice=Docs as architecture; Role=Catalog + definitions of the 13 repo-local agents — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `GEMINI.md`: Practice=Docs as architecture; Role=Harness-specific deltas; defers to `CLAUDE.md` — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `docs/high_signal_file_index.json`: Practice=Retrieval index; Role=491 curated entries with `how_to_run` / `risk_notes` — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `docs/llm-repo-index.json` + `.md`: Practice=Retrieval index; Role=758-file generated coverage map + human entrypoint — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `scripts/check-doc-indexes.mjs`: Practice=Retrieval index; Role=Validates 491 + 758 paths; CI-fails on dead paths — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] tam-MCP registries: Practice=Retrieval index; Role=Searchable skills / prompts / agents / URLs (capability findable) — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `server/src/telemetry/logger.js`: Practice=Logging; Role=`pino` structured JSON + scoped `logger.child({ scope })` — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `src/background/error-log.js`: Practice=Logging; Role=v1-schema, 1000-entry ring buffer; Sentry-lite + file sinks; Copilot remediation path — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `src/shared/error-log-client.js`, `src/content/shared/error-log-adapter.js`: Practice=Logging; Role=Shared client + content-script adapter — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `test/setup.js`: Practice=Testing; Role=`chrome.*` shims so extension logic runs under Node — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `mongodb-memory-server` (`MONGOMS_SKIP_IF_UNAVAILABLE`): Practice=Testing; Role=Server integration tests with air-gapped graceful skip — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `.github/workflows/{syntax-check,unit-tests,extension-smoke}.yml`: Practice=Testing; Role=Three CI gates over 707 tests — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `server/cli/call.js` · `server/src/routes/call.js` · `server/mcp/call-mcp-server.js`: Practice=Dual-mode; Role=Same six CallCard ops as CLI + HTTP + MCP over one core — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `127.0.0.1:8787` API + `packages/mcp-server/src/client.js`: Practice=Dual-mode; Role=One tested API core; extension and 13-tool MCP server as thin clients — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [parameter] `live-hub-toolkit/src/core` + `cli/*` + `local_fs_host` bridge: Practice=Dual-mode; Role=One engine, CLI + application surfaces; `node --test` exercises it — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#appendix-a-practice-artifact-inventory
- [definition] Building Codebases for Machine Collaborators — A June 2026 mdb-tam review shows how generated docs, retrieval indexes, logs, tests, and shared API surfaces help agents inspect and verify a codebase. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#building-codebases-for-machine-collaborators
- [definition] 1. The collaborator changed; the codebase must answer for it — A human engineer onboards once and accrues a mental model that survives across months. An LLM agent depends on the context its harness supplies on each call. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#1-the-collaborator-changed-the-codebase-must-answer-for-it
- [definition] 7. Test-centric design: dual-mode the capability across CLI, API, and application — This is the hinge, and it is where verifiability and legibility meet. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#7-test-centric-design-dual-mode-the-capability-across-cli-api-and-application
- [definition] 8. The practices as implemented — The six practices are wired into the running system, not aspirational. Each row maps a practice to what it buys a machine collaborator and to the mdb-tam mechanism that implements it; paths are repo-relative unless marked harness-level, and every path was confirmed to exist at the time of writing. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#8-the-practices-as-implemented
- [definition] 9. Implementation considerations — Six points, drawn from how this system is built. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#9-implementation-considerations
- [definition] 10. Conclusion — The collaborator changed. It depends on harness-supplied context, it reads the codebase a slice at a time, and its output is probabilistic — frequently right, occasionally confidently wrong, with no tonal tell between the two. — https://llms-explorer.com/blog/building-codebases-for-machine-collaborators/#10-conclusion
## TSE Strategy Backtest Scoreboard — Project Briefing
- [snippet] End-to-end run flow: datasets//cases.json (blind input: title + first message only) — `datasets//cases.json (blind input: title + first message only)` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#end-to-end-run-flow
- [snippet] Install (core): pip install -r requirements.txt # jsonschema>=4.26.0 — the only core dependency — `pip install -r requirements.txt # jsonschema>=4.26.0 — the only core dependency` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#install-core
- [snippet] Reproduce the seed scoreboard from a local Customer A export: _SRC=/path/to/customer-a-export python harness/seed_.py — `_SRC=/path/to/customer-a-export python harness/seed_.py` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#reproduce-the-seed-scoreboard-from-a-local-customer-a-export
- [snippet] Add your own strategy and score it: python harness/new_strategy.py my-method-v1 --title "My method" --kind hybrid — `python harness/new_strategy.py my-method-v1 --title "My method" --kind hybrid` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#add-your-own-strategy-and-score-it
- [snippet] Add your own strategy and score it: python harness/freeze.py strategies/my-method-v1 — `python harness/freeze.py strategies/my-method-v1` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#add-your-own-strategy-and-score-it
- [parameter] 1: Strategy=Pure MongoDB skill knowledge; n=244; Defensibility=100%; Acc / gradable=90.3%; Raw acc=72.5% — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#standings-on-2026-05-29-all
- [parameter] 2: Strategy=Documented flowchart corpus; n=244; Defensibility=86%; Acc / gradable=63.5%; Raw acc=51.0% — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#standings-on-2026-05-29-all
- [parameter] 3: Strategy=A colleague's Customer A flowchart bundle; n=244; Defensibility=37%; Acc / gradable=26.3%; Raw acc=21.1% — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#standings-on-2026-05-29-all
- [parameter] **"Which diagnosis method is best?" was anecdotal**: Frozen panels + deterministic scorecards + a ranked leaderboard make it measurable — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **A methodology grading its own exam**: The judge is a separate, versioned, methodology-blind artifact; predictor ≠ judge — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Ground-truth leakage into a "blind" test**: Strategies see only `initial_prompt`; resolutions live in a separate file the predictor never opens — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Post-hoc tampering to inflate a score**: Content-hash pinning plus a `validate.py` immutability pass reject edits to tested artifacts — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Stale or hand-edited scoreboards**: Every scorecard and the leaderboard are generated; CI fails if validation regenerates a tracked file — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Non-comparable scores**: Each run pins strategy + panel + judge + ground truth by hash; only same-panel, same-ground-truth runs are ranked together — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Accuracy hiding cost**: An efficiency block scores compute, navigation cost, size, and human-follow time — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Weak ground truth read as "accuracy"**: A `ground_truth_caveat` is stamped on every scorecard; scores are framed as plausibility until `r2` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Improving a method corrupting old results**: A new version is a new immutable folder; lineage tracked via `forked_from` — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Adding a contributor's method touching others' files**: Contribution is purely additive — new folders only; the scaffolder refuses to overwrite — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] Core harness (11 Python scripts): Path=`harness/*.py`; Approx. lines=~1,610 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] `/dr` deep-research runner (8 scripts): Path=`dr-harness/*.py`; Approx. lines=~650 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Artifact contracts: Path=`schemas/*.json` (9 files); Approx. lines=~360 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Scorer tests: Path=`tests/test_score.py`; Approx. lines=~110 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Documentation suite: Path=`docs/*.md` (17 files); Approx. lines=~1,440 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Post-run analyses: Path=`evaluations/*.md` (6 files: 5 analyses and an index); Approx. lines=~590 — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [definition] TSE Strategy Backtest Scoreboard — Project Briefing — A first-person briefing on a git-backed scoreboard that backtests case-diagnosis strategies blind against resolved support cases, with how to read its scores and add a strategy. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#tse-strategy-backtest-scoreboard-project-briefing
- [definition] Standings on 2026-05-29 *(all)* — Full corpus, panel `blind-244-v1` (calibrated rubric), ground truth `r1-autoclose-fallback`: — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#standings-on-2026-05-29-all
- [definition] Customer-data handling — The cases are real customer support cases (a Customer A export taken through TS Tools). The rule is to treat every case as customer-confidential and to store only the blind inputs and the resolutions grading requires — no fuller exports, PII, or internal URLs beyond that (`docs/SECURITY.md`). — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#customer-data-handling
- [definition] Blinding — Two-axis separation is built into the data model: blind inputs (`cases.json`) and ground truth (`resolutions-r1.json`) live in different files. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#blinding
- [definition] Secrets and network — The codebase reads no secrets — the only environment variable is `_SRC`, a local filesystem path used solely by the seeder. The scored ledger makes **zero in-process external calls**: no network, no database, no Model Context Protocol (MCP) calls. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#secrets-and-network
- [definition] Integrity model — The repo *is* the database. There is no server and no external state; integrity comes from canonical content hashes plus the `validate.py` CI gate, which is the read-back equivalent. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#integrity-model
- [definition] Run identity — `run_id = ______`. Every input is pinned by content hash, so a run is reproducible — and comparable to another run *only if* both share the same sample set **and** the same resolution version. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#run-identity
- [definition] Key modules — `common.py` (canonical hashing + JSON I/O), `new_strategy.py` (scaffold), `freeze.py` (write content hashes), `pin_runs.py` (stamp a run with its input hashes), `score.py` (grades → scorecard), `leaderboard.py` (runs → leaderboard), `validate.py` (the CI gate), and `seed_.py` (provenance… — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#key-modules
- [definition] Add your own strategy and score it — Contributing is additive: a new strategy is a new folder, and nobody edits another contributor's files (`CONTRIBUTING.md`). — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#add-your-own-strategy-and-score-it
- [definition] 9. Reading the Scores *(all)* — The standings rank strategies by plausibility, not verified accuracy. Two things explain why. — https://llms-explorer.com/blog/case-diagnosis-backtest-scoreboard-project-briefing/#9-reading-the-scores-all
## Case Study — Gamifying Incident-Response Training with a Rate-Limiting Fire Drill
- [parameter] Severity called correctly: What it measures=The IR role names severity 2; Target=within 10 min; Weight=blocking — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#23-the-rubric-what-the-drill-rewards
- [parameter] Rate-limiting diagnosed: What it measures=The IR role confirms a rate limit, not an outage; Target=within 20 min; Weight=major — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#23-the-rubric-what-the-drill-rewards
- [parameter] Mitigation suggested: What it measures=The IR role suggests a mitigation (backoff or a separate key); Target=within 30 min; Weight=major — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#23-the-rubric-what-the-drill-rewards
- [parameter] No unilateral quota promise: What it measures=Nobody promises a quota change; Target=none; Weight=blocking — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#23-the-rubric-what-the-drill-rewards
- [parameter] Fire-drill engine: Path=`src/background/firedrill-engine.js`; Role in this study=Lifecycle, drip, persona and complication injection, abort — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Drill state module: Path=`src/background/firedrill-state.js`; Role in this study=One of the two modules the engine imports — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Isolation safety test: Path=`tests/unit/firedrill-safety.test.js`; Role in this study=Checks drill modules and scenario files for forbidden imports and production hostnames — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Scorecard: Path=`src/background/firedrill-scorecard.js`; Role in this study=Observable types, weights, retro markdown — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Rate-limiting scenario: Path=`mcp-server/data/firedrill/scenarios/scenario-4.json`; Role in this study=Knowledge graph, drips, rubric, complications, persona — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Operator cards: Path=`docs/firedrill-ir-card.md`, `docs/firedrill-tam-ic-card.md`, `docs/firedrill-moderator-card.md`; Role in this study=JIMP roles under pressure — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [parameter] Program pitch: Path=`docs/firedrill-pitch.md`; Role in this study=Training framing and support-plan milestone — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [definition] Case Study — Gamifying Incident-Response Training with a Rate-Limiting Fire Drill — A training case study of an Atlas API rate-limiting fire drill: how its scorecard rewards restraint, and why team support multiplies, but cannot replace, individual expertise. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#case-study-gamifying-incident-response-training-with-a-rate-limiting-fire-drill
- [definition] §1 The challenge: readiness you cannot prove, and expertise you cannot clone — Before the fire-drill program, IR readiness was asserted, not demonstrated. The team had the JIMP, an escalation matrix and capable engineers. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#1-the-challenge-readiness-you-cannot-prove-and-expertise-you-cannot-clone
- [definition] §2 The solution: a gamified, isolated fire drill — The IR program built fire drills into the MDB Case Assistant extension (see the [project pitch](/blog/mdb-case-assistant-project-pitch/)). — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#2-the-solution-a-gamified-isolated-fire-drill
- [definition] §2.1 The engine and its safety model — The fire-drill engine (`src/background/firedrill-engine.js`) creates a simulated case (`DRILL-NNNN`) and drives it through a strict lifecycle: PREFLIGHT, RUNNING, CONCLUDED, with an ABORTED path always available. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#21-the-engine-and-its-safety-model
- [definition] §2.2 The rate-limiting scenario — Scenario 4 (`scenario-4.json`) is an Atlas API throttle rated severity 2 (S2) in the IR playbook. It is realistic, multi-step and ambiguous. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#22-the-rate-limiting-scenario
- [definition] §2.3 The rubric: what the drill rewards — The scorecard (`firedrill-scorecard.js`) is the heart of the design. It does not quiz individuals on encyclopedic Atlas knowledge. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#23-the-rubric-what-the-drill-rewards
- [definition] §2.4 The pedagogy underneath the game — I read the drill through established learning-science ideas: - **Productive failure** (Kapur): the scenario is hard enough that teams struggle before the retro consolidates the lesson, and the struggle helps the lesson stick. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#24-the-pedagogy-underneath-the-game
- [definition] §4 The result: what the drills taught — These observations are my summary of the drills (see Appendix B), not measured data, and the same pattern is visible in the rubric design itself. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#4-the-result-what-the-drills-taught
- [definition] §5 The conclusion — The fire drills converge on a two-part finding. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#5-the-conclusion
- [definition] Appendix A: Real artifacts referenced — Paths are relative to the root of the MDB Case Assistant repository, which is internal to MongoDB. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-a-real-artifacts-referenced
- [definition] Appendix B: Method and honesty note — This is a training case study, not a customer marketing asset. The mechanics described in §2 are real and cited above. — https://llms-explorer.com/blog/case-study-gamifying-incident-response-training/#appendix-b-method-and-honesty-note
## Conceptual vs proprietary llms files
- [parameter] file: source axis=`llms.txt`, `llms-full.txt`, `llms-facts.txt`; concept axis=`llms-concepts.txt`, `/t//llms-facts.txt`, concept packs — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [parameter] the authority: source axis=the publisher; concept axis=the concept — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [parameter] the unit of truth: source axis=a page; concept axis=a claim about the concept — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [parameter] what "correct" means: source axis=the link resolves and says what the description said; concept axis=the claim survives comparison with every other source's claim — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [parameter] what a disagreement is: source axis=possible between pages, versions, or claims from one publisher; concept axis=routine across sources under one concept — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [parameter] anyone (no account): may=read every file, every conflict record, the ladder; may not=submit — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [parameter] contributor (account): may=submit a unit with a source URL and anchor; it runs through `resolve`; may not=write a unit without a source; skip the ladder — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [parameter] maintainer: may=settle ties in the moderation queue with a note; reject a submission; may not=overwrite a ladder verdict without a note in the record — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [parameter] fork owner: may=keep a private tree with its own `precedence.json`; propose merges back as diffs of units; may not=push to the public tree directly — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [parameter] the lint: may=block any file with a High finding from being served; may not=be bypassed — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [definition] Conceptual vs proprietary llms files — A proposed governance design for conceptual llms files compares scoped claims, retains disagreements, and distinguishes source provenance from proof of correctness. — https://llms-explorer.com/blog/cllms-vs-proprietary/#conceptual-vs-proprietary-llms-files
- [definition] Two axes — This site organizes llms files along two axes. — https://llms-explorer.com/blog/cllms-vs-proprietary/#two-axes
- [definition] The most correct idea overwrites — On the concept axis, units can compete. The proposed `resolve` job first checks whether they answer the same question at the same scope: subject, version, date, platform, and population. — https://llms-explorer.com/blog/cllms-vs-proprietary/#the-most-correct-idea-overwrites
- [definition] The precedence ladder — After the scope and subject-matter fit checks, compare the rungs in order; the first difference decides, and a tie falls to the next. Source grade is a policy preference, not a guarantee of truth, so a standard’s definition cannot substitute for a measurement of adoption. — https://llms-explorer.com/blog/cllms-vs-proprietary/#the-precedence-ladder
- [definition] Disagreements stay visible — A conflict that the ladder settles produces a winner in `llms-facts.txt` and a loser in a `## Disagreements` section of the pack's `llms-full.txt`. A conflict it cannot settle puts both there. — https://llms-explorer.com/blog/cllms-vs-proprietary/#disagreements-stay-visible
- [definition] Governance — Who can overwrite what, on the public tree: — https://llms-explorer.com/blog/cllms-vs-proprietary/#governance
- [definition] Rights — What a conceptual file may contain is narrower than what a proprietary one may: - **Links** — the default way this design points readers to a publisher’s material. - **Facts in our words, with anchors** — yes. — https://llms-explorer.com/blog/cllms-vs-proprietary/#rights
- [definition] Honesty note — Three things the reader should know before trusting any of the above. — https://llms-explorer.com/blog/cllms-vs-proprietary/#honesty-note
## Comparing Output Quality Across Claude Model Tiers and Effort Levels
- [snippet] Appendix A — Reproducibility: T1: A water tank holds 240 liters and starts empty. Pipe A fills it at 8 liters — `T1: A water tank holds 240 liters and starts empty. Pipe A fills it at 8 liters` — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#appendix-a-reproducibility
- [parameter] Small: Model=Haiku 4.5; Model ID=`claude-haiku-4-5-20251001`; Role=Fastest / lowest-cost tier — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#21-models-under-test
- [parameter] Mid: Model=Sonnet 4.6; Model ID=`claude-sonnet-4-6`; Role=Balanced quality/cost — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#21-models-under-test
- [parameter] Large: Model=Opus 4.8; Model ID=`claude-opus-4-8`; Role=Most capable tier — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#21-models-under-test
- [parameter] Frontier: Model=Fable 5; Model ID=`claude-fable-5`; Role=**Not tested — returned "currently unavailable" this run** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#21-models-under-test
- [parameter] Haiku 4.5: T1 quant=10; T2 logic=10; T3 code=9¹; T4 writing=9²; Total / 40=**38** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#31-experiment-a-model-tier-neutral-prompting
- [parameter] Sonnet 4.6: T1 quant=10; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**40** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#31-experiment-a-model-tier-neutral-prompting
- [parameter] Opus 4.8: T1 quant=10; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**40** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#31-experiment-a-model-tier-neutral-prompting
- [parameter] Fable 5: T1 quant=—; T2 logic=—; T3 code=—; T4 writing=—; Total / 40=**n/a (unavailable)** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#31-experiment-a-model-tier-neutral-prompting
- [parameter] Haiku — low effort: T1 quant=**0**³; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**30** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#32-experiment-b-effort-level-haiku-opus
- [parameter] Haiku — high effort: T1 quant=10; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**40** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#32-experiment-b-effort-level-haiku-opus
- [parameter] Opus — low effort: T1 quant=**0**⁴; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**30** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#32-experiment-b-effort-level-haiku-opus
- [parameter] Opus — high effort: T1 quant=10; T2 logic=10; T3 code=10; T4 writing=10; Total / 40=**40** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#32-experiment-b-effort-level-haiku-opus
- [parameter] Haiku: Low=30; High=40; Effort Δ=**+10** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#33-interaction-rq3
- [parameter] Opus: Low=30; High=40; Effort Δ=**+10** — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#33-interaction-rq3
- [definition] Comparing Output Quality Across Claude Model Tiers and Effort Levels — A reported June Claude comparison with one response per condition; raw outputs are missing, and its score differences do not establish causal effects. — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#comparing-output-quality-across-claude-model-tiers-and-effort-levels
- [definition] 2.2 The "effort level" lever — what it is and is not — There is no public per-request "reasoning effort" dial exposed through the subagent interface I used, so **effort here is a prompt-induced proxy**, operationalized as two prompt regimes applied to the *same* benchmark: — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#22-the-effort-level-lever-what-it-is-and-is-not
- [definition] 3.4 Cross-experiment note — Neutral-prompt Haiku and Opus both scored T1 correctly (they were free to show work): Haiku's total (38) landed *between* the low- and high-effort conditions, while Opus's (40) matched the high-effort condition outright. — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#34-cross-experiment-note
- [definition] 6. Conclusion — The preserved tables report neutral scores of 38, 40, and 40 and a 10-point prompt-regime difference for Haiku and Opus, entirely in one arithmetic item. They do not establish a causal prompting advantage, the absence of an interaction, or a cheaper model-selection strategy. — https://llms-explorer.com/blog/comparing-output-quality-across-claude-model-tiers-and-effort-levels/#6-conclusion
## Two thousand artifacts, three taxonomies: crawling a customer engagement
- [snippet] Commands: /crawl-cust2llms "" --depth standard — `# one run per account; the pack is written into that account's own folder` — https://llms-explorer.com/blog/crawling-a-customer-engagement/#commands
- [snippet] Reproduce: xattr "/.gdoc" — `# 1. resolve one stub by hand first — if this prints nothing, nothing else will work` — https://llms-explorer.com/blog/crawling-a-customer-engagement/#reproduce
- [parameter] Artifacts enumerated: Account A=945; Account B=326; Account C=740 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#inputs
- [parameter] Google-native stubs: Account A=119; Account B=202; Account C=106 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#inputs
- [parameter] Taxonomy: Account A=`NN Name` (numbered, spaces); Account B=freeform; Account C=`NN_Name` (numbered, underscores) — https://llms-explorer.com/blog/crawling-a-customer-engagement/#inputs
- [parameter] Has a `_meta/` self-map: Account A=yes; Account B=no; Account C=no — https://llms-explorer.com/blog/crawling-a-customer-engagement/#inputs
- [parameter] Stubs resolved: A=116 / 119; B=201 / 202; C=99 / 106 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Deep-read: A=47; B=18; C=16 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Index documents examined: A=23; B=14; C=12 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Atomic tagged facts: A=61; B=45; C=51 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Claims tagged stale: A=9; B=3; C=2 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Conflicts recorded: A=6; B=5; C=4 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] Redactions: A=1; B=0; C=0 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] critical: A=23; B=16; C=9 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] high: A=48; B=54; C=67 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] normal: A=152; B=220; C=160 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [parameter] peripheral: A=722; B=36; C=504 — https://llms-explorer.com/blog/crawling-a-customer-engagement/#outputs
- [definition] Two thousand artifacts, three taxonomies: crawling a customer engagement — Three anonymized engagement crawls report 2,011 artifacts and 416 resolved Drive stubs, while exposing context-selection and confidentiality-check limits. — https://llms-explorer.com/blog/crawling-a-customer-engagement/#two-thousand-artifacts-three-taxonomies-crawling-a-customer-engagement
## Customer Dashboard — Project Briefing
- [snippet] C4 Level 1 — System context: Operator — `Operator` — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#c4-level-1-system-context
- [parameter] **Context fragmentation** — preparing for a customer call means opening 6–10 tabs across Hub, Salesforce, Slack, monday.com, Jira, Aha!, Drive, and Glean: Aggregates all sources into a single indexed corpus; generates a pre-call context report in under a minute — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Account context lives in two systems** — the technical story is in Hub, but ownership, annual recurring revenue (ARR), and roadmap status live in Salesforce and Aha!: Content scripts extract Salesforce Account fields and Aha! item status directly into the same corpus, so a brief reflects both — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Case status blind spots** — severity escalations are only visible if you happen to be looking at Hub: Polls open cases in the background and surfaces Chrome notifications on severity or status changes — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **monday.com maintenance overhead** — board items go stale; reconciling them against Hub cases is manual: Automated monday.com reconciliation: creates new items, archives resolved cases, updates rows via LLM-assisted diffing — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Meeting context loss** — notes from Granola, Plaud recordings, and Slack threads are siloed: Ingests meeting transcripts and recordings through native host bridges and content scripts, and indexes them into the local corpus — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Report generation time** — writing a pre-call summary or case analysis by hand takes 30–60 minutes: LLM report generation with configurable prompts produces structured reports from the indexed corpus in under a minute — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Atlas diagnostics access** — diagnostics snapshots require navigating to each project separately: Per-cluster diagnostics snapshot refresh and storage directly from the extension dashboard — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **No unified to-do surface** — action items from calls, Slack, and cases live in different places: Floating always-on-top to-do window with keyboard shortcut — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Slow onboarding** — new TAMs rebuild account history from scratch: A new TAM's own dashboard rebuilds account history from the sources automatically instead of starting from scratch — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] **Slow escalation response** — priority changes surface only if someone is watching: Case-status notifications can arrive before the customer sends a follow-up — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#3-problems-solved-leadership-team
- [parameter] Chrome extension (background, content, UI): Approx. lines=~102,000; Tests=294 (Vitest) — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Local backend server (`server/`): Approx. lines=~18,000; Tests=346 (Vitest) — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Live Hub Toolkit (`live-hub-toolkit/`, which generates customer hub documents): Approx. lines=~3,900; Tests=30 (`node --test`) — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Native host bridges (`native-host/`): Approx. lines=~6,900; Tests=Python AST + integration — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Account Context MCP server: Approx. lines=~800; Tests=37 (`node --test`) — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Documentation suite: Approx. lines=78 markdown files; Tests=— — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] **Total**: Approx. lines=**~131,600**; Tests=**707** — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#4-scope-of-work-leadership-reviewers
- [parameter] Local backend isolation: Server binds to `127.0.0.1:8787` by default, so it is not reachable from outside the machine — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#network-and-backend-security
- [parameter] Origin allow-list: All state-changing backend requests checked against an explicit origin allow-list in `server/src/middleware/origin-check.js`; cross-origin requests are rejected — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#network-and-backend-security
- [parameter] Timing-safe token verification: Backend bearer token comparison uses `crypto.timingSafeEqual`; not `===` (`server/src/middleware/auth.js`) — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#network-and-backend-security
- [parameter] No query-string tokens: Auth sent as `Authorization: Bearer` — never embedded in URLs — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#network-and-backend-security
- [parameter] Service worker: Path=`src/background/service-worker.js`; Owns=message routing, alarms, sync engines, corpus writes; Talks via=`chrome.runtime.sendMessage`, `chrome.alarms`, IndexedDB, native messaging, loopback HTTP — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#c4-level-2-chrome-side-containers
- [parameter] Options page: Path=`src/options/`; Owns=settings and credentials UX; Talks via=runtime messages + `chrome.storage.*` — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#c4-level-2-chrome-side-containers
- [parameter] Popup: Path=`src/popup/`; Owns=quick actions and dashboard launch; Talks via=runtime messages + `chrome.storage.*` — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#c4-level-2-chrome-side-containers
- [parameter] Dashboard pages: Path=`src/dashboard/`; Owns=main operator workspace, overlays, to-do tooling; Talks via=runtime messages + `chrome.storage.*` — https://llms-explorer.com/blog/customer-dashboard-project-briefing/#c4-level-2-chrome-side-containers
- [parameter] Offscreen document: Path=`src/offscreen/`; Owns=SSE client, LLM streaming (kept alive by silent `