64x-lunicorn A claim is only as good as the source that owns it. Memory and secondary write-ups age silently, and an unchecked fact in research becomes a requirement in the spec.
1. Isolate the claims
Write each claim as one checkable sentence, with the version or date it depends on ("SQLite in WAL mode allows concurrent writers", not "SQLite handles concurrency").
Done when every claim is one sentence that a single source could confirm or contradict.
2. Find the owning source
The primary source is whoever owns the fact:
- Technical: official documentation, the specification, the source code, the changelog or release notes of the exact version in use. Prefer the Context7 docs server for libraries, fetch the page otherwise.
- Domain: the regulator or legal text, the standards body, the company's own publication, first-party data.
Blog posts, forum answers and AI summaries only lead to the primary source; they are never cited as the source, because they are where claims drift.
Done when each claim has a primary source, or the search for one is recorded as failed.
3. Delegate long reading
When the claims need many sources or long documents and the discussion can go on meanwhile, hand the list to a background agent with steps 1, 2 and 4 of this skill and the target sources.md. Keep working; merge its entries when it reports.
Done when every claim is either checked or assigned to a running agent.
4. Record
In a research object, add one row per claim to its sources.md:
| Claim | Source | Accessed | Version | Confidence |
- high: the primary source states it directly for this version.
- medium: the primary source implies it, or states it for a nearby version.
- low: no primary source found; this is an assumption.
Outside a research object, give the same fields inline with the answer.
Done when every checked claim has a row with all five fields.
5. Report back
Group the claims as confirmed, contradicted and unverified. Say plainly when a claim that the discussion relies on was contradicted, and keep unverified claims labelled as assumptions wherever they are used, so they are never promoted to fact by repetition.
Done when each claim carries one of the three labels in the answer.