deep-research

Files (1)

• SKILL.md

SKILL.md (first lines)

# Deep research

A brief is a set of claims someone will act on, so it obeys evidence rules.

1. **Decompose first.** Write the three to six sub-questions whose answers
   decide the verdict, and search for those — not for the topic. A search
   plan in your notes beats ten reactive queries.
2. **Search finds, fetch reads.** Use `search_web` to locate candidate
   pages, then `fetch_page` on the ones worth reading. Prefer primary
   sources (the company, the paper, the regulator) over aggregators
   repeating them; an aggregator citing a primary is a pointer, not a
   second source.
3. **Two independent sources for every load-bearing claim.** Independent
   means separately produced — two articles quoting one press release are
   one source. A claim with one source is a lead and is labelled as one.
4. **Cite per claim, with the date.** Beside each claim: `[url, fetched
   YYYY-MM-DD]` — the store's own provenance shape, extended to the web.
   A citation at the foot of the page attached to nothing is not one.
5. **Disagreement is a finding.** When sources conflict, report both
   numbers, name which you trust and why (recency, proximity, method) —
   never a silent average, never the friendlier figure.
6. **Verdict first.** The brief opens with the answer in two or three
   sentences, then the evidence, then the detail. A reader who stops after
   the first paragraph should leave correctly informed.
7. **Close with Gaps.** What you looked for and could not find, what is
   behind paywalls or dated, and what a longer look would check — as
   plainly as the findings. A brief that hides its edges is wrong even
   where its claims are right.
8. **Never pad.** If the honest answer is thin, deliver it thin and say
   why. Model knowledge may frame a question but never settles a claim —
   anything load-bearing was fetched, or it is labelled as unverified.

From briant92/Agentlings, Apache-2.0.

More from briant92/Agentlings