Context files and concept facts files: what an agent loads

Two things on this site are meant to be loaded whole into a context window rather than browsed. Both are listed on /context/ for a person and in /context.md for an agent, and both are served as text/markdown with the describedby link back to /llms.txt.

1. Context files

A context file is one of the hub’s own research reports: a dated, sourced write-up of one subject, produced by the research skills this site publishes (/dr, rabbithole, concept-family-explorer). Each is mirrored twice:

They exist on this site so that a fact in a concept pack can name a source that resolves publicly instead of a path inside a private repository. They are not twinned into the site’s own llms-full.txt: three hundred reports would drown the family this build curates, so the index lists them under /context.md and leaves each one a single fetch away.

2. Concept facts files

A concept facts file is one concept’s pack as a markdown list — the same facets and facts the node page under /tree/ renders, written to /downloads/concepts/<slug>.md by gen_concept_facts.py on every build. The grammar:

<!-- llms-explorer concept facts · <page URL> · pack <date> · ~<N> tokens -->
# <Concept>

> <summary>

Parent: [<parent>](<parent page>) · <facets> facets · <facts> facts · page: <page URL>

## <Facet title>
- <fact> — [source](<URL>)
  - <sub-point> — [source](<URL>) *(<note>)*

## Related concepts
- [<concept>](<page>) — <relation>

## Context files
- [<title>](/downloads/sources/<hub>/<name>.md)

Every fact keeps the source it was extracted with. A source that is a URL is a link; a source that is a local path (the earliest packs cite files under a skills directory) is shown in code and never linked, because a link to it resolves nowhere. Sub-points nest under the point that introduces them, from the level the pack builder derived from the source’s own list indentation. The token estimate in the banner is the family’s declared estimator (bytes ÷ 4), placed in the file because downloads carry no per-file X-Markdown-Tokens header — see usage §2 for what a twin and a download each promise.

3. How a row is filed

/context/ groups both shapes under the roots of the concept tree:

• a concept facts file goes under its own root, walked up parent_slug in tree.json;
• a context file goes under the root of the concept that cites it most; a file cited from two subtrees is listed under the heavier one and links every concept that cites it;
• anything the tree cannot place — a pack whose node is not in this snapshot, a report nothing cites yet — lands under a synthetic last root, Not yet filed, so it is visible rather than absent.

context.json is generated from the committed sources, packs and tree, never from the clock: its generated stamp is the tree’s, and CI diffs the committed file against a fresh run.

4. Reading it the one-hop way

1. Fetch /context.md. It lists every context file and every concept facts file with its absolute URL, grouped by root, in one file.
2. Search it for your term. A hit names the file.
3. Fetch that file. Stop.

When the term is a concept rather than a subject, the root /llms.txt already lists every /tree/<slug>/ page; the facts file is at /downloads/concepts/<slug>.md with the same slug.

5. Rights

Context files are the hub’s own words about sources it read, with those sources cited. No third-party full text is republished on either surface; the mirrored llms-full.txt files the directory grades are scored, not served. The ethos page carries the whole position.