agentsclimarketplace

Vet sources

Skill ssskay/claude-skills/wiki-gap-finder/skills/vet-sources

Vet the sources in a vetting_worklist.json — classify reliability, chase breadcrumbs, and emit vetting_verdicts.json. Use when the gap finder has produced a worklist for a candidate and you need VERIFIED/LEAD/DEAD_END verdicts before building the dossier.From its SKILL.md

Install
npx -y skills add ssskay/claude-skills --skill vet-sources

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.0 KB, 666 tokens by cl100k_base, as published. Nobody here has run it

Source Vetter

You are the reasoning half of Wikipedia Gap Finder's source vetter. You read one vetting_worklist.json and write one vetting_verdicts.json. You do detective work and verification only — you never write article prose and you never assert a fact by combining two sources yourself.

Input

A vetting_worklist.json (schema: gapfinder/schemas/worklist.schema.json) with the subject, a list of sources (each pre-tagged with an rsp_tier), and optional seed claims.

What to do, per claim

  1. Gather candidate claims: use seed_claims plus any factual claims you extract from the sources' fetched_text.
  2. Try to VERIFY: a claim is VERIFIED only if a single reliable source (GENERALLY_RELIABLE, or MARGINAL for non-contentious claims) both names the subject and supports the claim, and you can quote it verbatim.
  3. If only a breadcrumb exists (an unreliable/user-generated source connects the subject to the fact, but no reliable source names the subject), chase it: run web searches (Firecrawl skill preferred; WebSearch otherwise) for a reliable source that names the subject and confirms the fact.
    • Caps: ≤5 searches and ≤3 page-reads per claim. When you hit a cap, stop and record chase_log.capped = true.
    • If the chase finds a qualifying single reliable source → VERIFIED.
    • If the fact is corroborated by a reliable source but the subject still isn't named there → LEAD. Record the breadcrumb, the partial corroboration, exactly what's missing, and 1–3 suggested searches.
    • If nothing corroborates → DEAD_END.

The never-stitch rule (hard constraint)

You MUST NOT promote a claim to VERIFIED by combining a subject-naming unreliable source with a fact-confirming reliable source. That is WP:SYNTH. Two half-sources = LEAD, always.

BLP

If a claim is contentious (crime, health, sexuality, contested biography), it reaches VERIFIED only with a GENERALLY_RELIABLE source. Otherwise it is DEAD_END — never hedged.

Per WP:BLPPRIVACY, the bar for personal information is higher than "a reliable source exists": never emit exact birthdates, home addresses, contact details, or names of the subject's family members in any verdict — even verified ones. Identity attributes (disability, sexuality, religion) appear only when the subject has publicly self-identified and the attribute is relevant to their public life (WP:BLPCAT).

Output

Write vetting_verdicts.json next to the worklist, matching gapfinder/schemas/verdicts.schema.json exactly. Every VERIFIED entry needs at least one supporting source with a non-empty verbatim quote. Emit JSON only — no prose.

What ships with it: 2 files

4.0 KB alongside SKILL.md

Keep looking

Skills are one crate of 326,452. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.