Research
Never invent sources, numbers or quotes. Unknown stays unknown.
Size it
- Quick (one fact or comparison): 2–4 searches, primary sources first, an answer with links.
- Deep (the topic needs a report): the steps below. In Claude, if the
anthropic-skills:deep-researchskill is available and the user wants a long narrative report, use it; this skill's steps still apply to the evidence standard. - Evaluate a library or repo: inspect the actual repo (license, last commit, open issues, install scripts, hooks, telemetry, network calls, size) before recommending it. Record the URL, commit and license. Never install or run it to "try it" without approval.
- Data or experiment check: reproduce the numbers from the raw data with a script before interpreting them; state sample sizes and what could confound the result.
Deep research
- Frame: the question, why it matters, what decision it feeds, and 3–7 sub-questions.
- Gather: search each sub-question (fan out to parallel subagents when available, one sub-question each). Prefer primary sources: docs, source code, papers, changelogs, official data.
- Cross-check: every important claim needs two independent sources or one primary one. Note dates; flag anything older than a year in a fast-moving field.
- Contrarian pass: look for the strongest evidence against the emerging conclusion.
- Synthesize: the answer first, then findings with inline citations, confidence
(high/medium/low) per finding, open questions, and the sources list. Keep the files in
research/<slug>/in the project when the user wants to keep them.