Worked example: the Anthropic hiring concept pack

Every recipe on this site shows a retrieval technique against a hypothetical corpus. This one shows the corpus getting built — an honest walkthrough of one real concept pack in this repo, not a marketing pitch for the pipeline that made it.

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? The subject picked was concrete and checkable: “what does Anthropic’s hiring process actually look like, end to end?” The output lives at the repo root as llms-anthropic-hiring.txt and llms-facts-anthropic-hiring.txt, indexed from llms.txt.

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. Six independent sources fed the pack: FinalRoundAI’s interview guide, Glassdoor and TeamBlind candidate reports, levels.fyi for compensation, Anthropic’s own published candidate-AI-usage guidance, and a candidate’s substack account of a recent loop. Verified 2026-09-07 — the date stamped into both files, so a reader can judge staleness without re-running anything.

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 candidate. llms-facts-anthropic-hiring.txt is meant to be grepped and cited: one claim per line, each ending in a [src:] tag. Splitting them means a single fact can be corrected, re-sourced, or downgraded without touching the narrative that explains it — and an agent that only needs the citation behind one number never has to load the whole pack to find it.

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. Here that’s Anthropic’s own candidate guidance, FinalRoundAI’s guide, and 200+ Glassdoor/TeamBlind candidate reports converging on the same rounds-and-timeline picture:

Anthropic interview process has 4-6 rounds and takes 3-6 weeks total (SWE: 3-5 weeks, Research: 4-7 weeks) [src: finalroundai.com, glassdoor.com]

That rating describes the pack as a whole, not every line inside it. Individual facts in llms-facts-anthropic-hiring.txt carry whatever [src:] tags reflect their own actual sourcing — including lines resting on a single source, like the claim that shapes the whole pack’s emphasis:

Values and mission alignment interview is weighted equally with all technical rounds combined [src: finalroundai.com]

Compensation figures are quoted as sourced examples, not restated as a claim of this site’s own, because they carry that same one-source-plus-one-aggregator trail rather than the three-way agreement behind the pack’s overall rating:

Software Engineer (L4/L5) total compensation range: $250K-$400K, with base $180K-$230K [src: finalroundai.com, levels.fyi]

And where only one source exists at all, the fact still ships, attributed honestly rather than dropped or folded into the high-confidence average:

The culture interview reveals “how people think” and is where confident, experienced candidates often fail [src: ridhimakhurana.substack.com]

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. An agent with a hiring question loads the concept pack’s 10-line index next, jumps straight to the matching topic section (e.g. “Compensation” for a negotiation question), and only opens llms-facts-anthropic-hiring.txt when it needs the exact source behind a specific number. Most questions never need that last hop.

What to copy for your own subject

Name the subject concretely enough to check. Run /dr per topic-shaped question — if the subject is broad enough that you don’t yet know which topics matter, run concept-family-explorer first to map the family and decide what’s worth researching before handing topics to /dr. Write every claim as one line with a [src:] tag in a facts file, separate from the narrative pack. Register both under one entry in llms.txt with a confidence rating you can defend by counting sources out loud — not a number that just feels right.