agentleFS
Nightflight6Nightflight6/open-agent-skills/skills/en/grounded-claims-reviewer/SKILL.md

grounded-claims-reviewer

Skill1 starsChanged 4 months agoMITView on GitHub

Extract, classify, prioritize, verify, and audit material claims against available evidence while preserving uncertainty, provenance, scope, and source quality. Use for factual review of articles or reports, claim extraction, fact and evidence checking, technical or numerical verification, market or competitor claim review, unsupported-certainty detection, claim-inflation detection, and audits of Claim Ledgers or cited-source support. Do not use as the primary skill for ordinary copyediting, non-factual rewriting, writing from scratch, SEO or GEO planning, deep search-intent or generic topic research, a simple one-fact lookup, translation, summarization only, keyword research, or title generation.

The file

Show its rules and text

We split this skill into the rules it gives an agent

19 rules. About 0 are common, 13 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never fabricate a source, URL, citation, quotation, statistic, or item of evidence.

    line 18
  2. Distinctive:

    Never treat a source's existence as proof that it supports the exact claim.

    line 19
  3. Distinctive:

    Never search only for support for the user's preferred conclusion or hide material contradictory evidence.

    line 20
  4. Very specific:

    Never silently convert attributed, inferred, or unverified to verified.

    line 21
  5. Distinctive:

    Do not force citations onto every low-risk sentence, use mandatory sentence-by-sentence review, or create numerical confidence or universal …

    line 22
  6. Distinctive:

    Do not rewrite the full article unless a limited correction is necessary for the verification task.

    line 24
  7. Distinctive:

    Use only the least expansive mode that satisfies the request:

    line 29
  8. Distinctive:

    Follow an explicit mode or output requirement first.

    line 35
  9. Very specific:

    Read references/output-schema.md.

    line 42
  10. Very specific:

    Read references/claim-taxonomy.md.

    line 54
  11. Distinctive:

    Read references/evidence-quality.md.

    line 61
  12. Distinctive:

    Read references/verification-workflow.md and apply the risk-proportionate workflow.

    line 68
  13. Very specific:

    Use only keep, attribute, qualify, revise, remove, or verify as recommended_action.

    line 90
  14. Distinctive:

    Read references/claim-inflation.md to detect certainty, scope, causal, trend, attribution, forecast, and consensus inflation.

    line 95
  15. Distinctive:

    Read references/source-conflicts.md when reliable sources differ.

    line 97
  16. Very specific:

    If external access is prohibited or unavailable, do not fabricate.

    line 112
  17. Very specific:

    When revised_claim exists, downstream editors should use it without increasing certainty or removing scope.

    line 117
  18. Distinctive:

    Follow explicit user output requirements first.

    line 124
  19. Distinctive:

    Keep evidence references traceable and unresolved uncertainty visible.

    line 130

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Grounded Claims Reviewer

<!-- parity: purpose -->

Purpose

Extract, classify, prioritize, verify, and review material claims against available evidence. Preserve uncertainty, provenance, scope, time, and source quality while preventing unsupported certainty, fabricated evidence or citations, citation laundering, confirmation bias, and claim inflation.

Golden rule: Verification seeks truth, not support for a preferred conclusion.

<!-- parity: guardrails -->

Non-negotiable guardrails

  • Never fabricate a source, URL, citation, quotation, statistic, or item of evidence.
  • Never treat a source's existence as proof that it supports the exact claim.
  • Never search only for support for the user's preferred conclusion or hide material contradictory evidence.
  • Never silently convert attributed, inferred, or unverified to verified.
  • Do not force citations onto every low-risk sentence, use mandatory sentence-by-sentence review, or create numerical confidence or universal credibility scores.
  • Preserve meaningful uncertainty. Do not overstate a claim merely to make the output decisive.
  • Do not rewrite the full article unless a limited correction is necessary for the verification task.
<!-- parity: mode-routing -->

Route the mode

Use only the least expansive mode that satisfies the request:

  1. scan: quickly identify material and high-risk claims, obvious unsupported certainty, and the next claims to verify. Do not perform exhaustive external verification unless required.
  2. verify: perform substantive claim verification: extract, prioritize, research, compare evidence, assign status, and recommend correction.
  3. audit: review an existing Claim Ledger, evidence set, GEO Claim Ledger, verification table, or structured research output without rebuilding it unnecessarily.

Follow an explicit mode or output requirement first. Ask only when ambiguity would materially change the allowed work.

<!-- parity: inputs-interoperability -->

Establish inputs and preserve interoperability

Determine the content or ledger under review, source context, available evidence, audience, market, relevant time, and mode. Accept upstream Evidence Tables, SERP Observation Tables, Claim Ledgers, Answer Maps, GEO Content Briefs, SEO Content Briefs, user-provided documents, and official references.

Read references/output-schema.md. For a GEO Claim Ledger, preserve claim_id when practical, map claim to claim_text and evidence_ref to evidence_refs, retain provenance fields, then add reviewer fields. Never treat an upstream record as automatically verified.

<!-- parity: claim-selection -->

Select material claims

Identify candidate claims, then separate verifiable propositions from pure opinion, rhetoric, recommendations, and stylistic language. Opinion is not automatically a factual claim, but factual premises inside an opinion may require verification.

Prioritize by the consequence of error for reader understanding, technical correctness, business or product decisions, brand credibility, legal or compliance exposure, market or financial interpretation, safety, and publication reliability. Do not treat every sentence as equally important.

<!-- parity: taxonomy-materiality -->

Classify type, materiality, and time sensitivity

Read references/claim-taxonomy.md. Assign the narrowest applicable claim_type.

Use only High, Medium, or Low for materiality, evidence_quality, and time_sensitivity; these are qualitative judgments, not numerical scores. Current product capability, price, software version, regulation, policy, market data, management, ranking, availability, platform policy, and current AI functionality commonly have high time sensitivity. Do not impose unnecessary recency on stable facts.

<!-- parity: evidence-quality -->

Assess evidence quality

Read references/evidence-quality.md. Evaluate source authority, primary versus secondary status, directness, exact claim support, date, methodology, independence, context completeness, scope fit, and version or product fit.

A famous domain, professional appearance, or repeated unsupported claim does not make evidence High quality. Community evidence may verify that users reported something; it does not automatically establish that the experience is generally true.

<!-- parity: verification-workflow -->

Verify without confirmation bias

Read references/verification-workflow.md and apply the risk-proportionate workflow. If asked to “find evidence proving X,” evaluate whether reliable evidence supports X. Inspect material supporting and contradictory evidence when appropriate without creating false balance where evidence overwhelmingly favors one conclusion.

Reuse supplied evidence when appropriate, but compare it to the exact proposition before assigning a status.

<!-- parity: entailment-provenance -->

Check entailment and provenance

Match the exact proposition, object or product, version, time, geography, units, measurement definition, conditions, and scope. A source saying “supports up to 64 GB” does not verify “the default configuration is 64 GB.”

Trace material claims toward original evidence when practical. Citation chains do not increase strength by themselves. If a source only verifies that another source made a statement, use attributed; if support remains insufficient, use unverified.

<!-- parity: status-actions -->

Assign canonical status and action

Use only:

  • verified: reliable evidence directly supports the claim within its stated scope;
  • attributed: the source made the claim, but the proposition is not independently established;
  • inferred: a reasonable conclusion derived from evidence but not directly stated;
  • unverified: reliable support is insufficient;
  • contradicted: reliable evidence materially conflicts with the claim.

Use only keep, attribute, qualify, revise, remove, or verify as recommended_action. Produce the narrowest useful revised_claim when qualification or correction helps.

<!-- parity: inflation-conflicts -->

Detect inflation and source conflicts

Read references/claim-inflation.md to detect certainty, scope, causal, trend, attribution, forecast, and consensus inflation. Recommend only wording the evidence supports.

Read references/source-conflicts.md when reliable sources differ. Do not average conflicting facts, silently choose a preferred source, assume newest is always correct, or assume an official source resolves every methodological issue. Report unresolved material conflict.

<!-- parity: high-risk-claim-types -->

Handle high-risk claim types

  • For comparative and superlative claims, require a defined market, period, comparison universe, measurement criterion, source, and methodology. Otherwise qualify, verify, or remove.
  • For causal claims, distinguish association, mechanism, and causation. Do not infer causation from correlation.
  • For forecasts, preserve the forecast source, date, and horizon. Never rewrite an expected future value as a current fact.
  • For attribution, verify who said what without silently verifying the underlying proposition.
<!-- parity: external-tools -->

Use external tools within host policy

Remain provider-neutral and follow the host Agent or platform's tool, network, access, privacy, and approval policies. When current or external facts materially affect correctness, tools are available, and the user has not prohibited external access, verify them.

If external access is prohibited or unavailable, do not fabricate. Preserve unverified, attributed, or inferred as appropriate and state the limitation.

<!-- parity: publication-integration -->

Integrate reviewed wording safely

When revised_claim exists, downstream editors should use it without increasing certainty or removing scope. Keep corrections localized unless broader wording is necessary to prevent factual inconsistency.

Route broader editorial work to publication-ready-article-editor; do not turn claim verification into a full content rewrite.

<!-- parity: delivery -->

Deliver by mode

Follow explicit user output requirements first. Do not force a large table for a lightweight task. Otherwise return:

  • scan: High-risk claims, key concerns, and recommended next checks;
  • verify: verification summary, Claim Ledger, critical corrections, and remaining verification;
  • audit: audit findings, claim status changes, evidence gaps, and recommended actions.

Keep evidence references traceable and unresolved uncertainty visible.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.