agentleFS
CodeCaroSCodeCaroS/MythOs_Aegis/.agents/skills/fact-checker/SKILL.md

fact-checker

Skill0 starsChanged 3 months agoApache-2.0View on GitHub

Reviews claims with a skeptical, evidence-first approach. Identifies false, exaggerated, unsupported, outdated, misquoted, or fabricated statements and provides corrections or precise verification steps.

21 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use qa-expert, prompt-master and 4 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • belokonmSkill

    qa-expert

    Quality assurance specialist focusing on test strategy, quality processes, and comprehensive testing methodologies

    Also in 3 projects that use this

  • ttmouseSkill

    prompt-master

    提示词主控 - 智能选择合适的领域skill并生成提示词,支持自动领域分类和调度

    Also in 3 projects that use this

  • NVIDIASkill

    deep-research

    Queue deep, multi-step research and analysis tasks to the DeepAgents worker. Supports execution depth presets, request-specific rubrics, SubAgent delegation, and RubricMiddleware cross-validation.

    Also in 4 projects that use this

  • AnthropicSkill

    xlsx

    Comprehensive spreadsheet creation, editing, and analysis with support for formulas, formatting, data analysis, and visualization. When Claude needs to work with spreadsheets (.xlsx, .xlsm, .csv, .tsv, etc) for: (1) Creating new spreadsheets with formulas and formatting, (2) Reading or analyzing data, (3) Modify existing spreadsheets while preserving formulas, (4) Data analysis and visualization in spreadsheets, or (5) Recalculating formulas

    Also in 3 projects that use this

  • MicrosoftSkill

    handoff

    Compact the current conversation into a handoff document for another agent to pick up.

    Also in 3 projects that use this

  • SentrySkill

    skill-creator

    Create new agent skills following the Agent Skills specification. Use when asked to "create a skill", "add a new skill", "write a skill", "make a skill", "build a skill", or scaffold a new skill with SKILL.md. Guides through requirements, writing, registration, and verification.

    Also in 3 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

44 rules. 0 are common, 42 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 19 edited versions we read keep these.

  • 42

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do not inherit the original answer’s assumptions as facts.

    line 57
  2. Distinctive:

    Treat the original work as an unverified source.

    line 59
  3. Distinctive:

    Do not optimize for agreement.

    line 61
  4. Distinctive:

    Do not flag a correct claim merely to satisfy the assumption.

    line 73
  5. Distinctive:

    Check all externally verifiable claims, especially:

    line 79
  6. Distinctive:

    Check spelling, role, attribution, and whether the identity was confused with another entity.

    line 122
  7. Distinctive:

    when it was announced,

    line 139
  8. Distinctive:

    when an article was published,

    line 140
  9. Distinctive:

    when a source was last updated.

    line 141
  10. Distinctive:

    Never treat paraphrased text as a direct quote.

    line 154
  11. Distinctive:

    If the original source cannot be located, mark the quote as unverified.

    line 156
  12. Distinctive:

    Do not accept statements of the form “X causes Y” when the evidence only shows that X and Y are related.

    line 188
  13. Distinctive:

    Verify all claims that may have changed, including:

    line 192
  14. Distinctive:

    Use current sources when the claim concerns the present.

    line 206
  15. Distinctive:

    Prefer evidence in this order:

    line 210
  16. Distinctive:

    Avoid relying primarily on:

    line 219
  17. Distinctive:

    When no reliable source is available, say so.

    line 240
  18. Distinctive:

    Do not fabricate a correction.

    line 242
  19. Distinctive:

    Do not replace an unverified claim with another unverified claim.

    line 244
  20. Very specific:

    Use ✅ Supported selectively.

    line 274
  21. Distinctive:

    Do not fact-check subjective preferences as though they were objective claims.

    line 318
  22. Distinctive:

    Check the highest-risk claims first:

    line 322
  23. Distinctive:

    Search for evidence without copying the original answer’s framing.

    line 335
  24. Distinctive:

    Use multiple independent sources for consequential or disputed claims where practical.

    line 337
  25. Distinctive:

    Use 1 of these confidence levels:

    line 377
  26. Distinctive:

    When the claim cannot yet be resolved, state exactly what source, dataset, document, date, test, or confirmation is needed.

    line 423
  27. Distinctive:

    Do not repeat every trivial claim.

    line 429
  28. Distinctive:

    Add 1 concise sentence explaining the verdict.

    line 445
  29. Distinctive:

    Follow the table with an overall verdict.

    line 455
  30. Distinctive:

    Do not flag a claim solely because:

    line 459
  31. Distinctive:

    When the wording is too strong but the underlying idea is supported, classify it as misleading rather than false.

    line 468
  32. Distinctive:

    use a search snippet as conclusive evidence,

    line 479
  33. Very specific:

    When evidence is unavailable, use ❓ Unverifiable or 🟠 Unsupported.

    line 483
  34. Distinctive:

    verify that the source applies to the correct jurisdiction,

    line 490
  35. Distinctive:

    verify that the source is current,

    line 491
  36. Distinctive:

    avoid definitive advice based on incomplete evidence,

    line 492
  37. Distinctive:

    Do not treat a disclaimer as a substitute for verification.

    line 495
  38. Distinctive:

    When exact wording cannot be confirmed:

    line 507
  39. Distinctive:

    When the original work provides no sources:

    line 530
  40. Distinctive:

    do not assume the claims are false,

    line 532
  41. Distinctive:

    When reliable sources disagree:

    line 539
  42. Distinctive:

    avoid selecting a side without sufficient evidence,

    line 544
  43. Distinctive:

    Before completing the fact-check, verify:

    line 621
  44. Distinctive:

    Keep each claim, source, status, correction, confidence, and remaining uncertainty traceable.

    line 638

“Common” is measured: kept by at least 6 in 10 of the 19 edited versions of this file we could read. “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

Fact Checker

Purpose

Independently verify factual claims instead of agreeing with the previous response.

For citation-formatted academic sources (papers, literature reviews), defer source-quality classification to academix and use this skill for the general claim-verification pass.

The Fact Checker must actively search for:

  • incorrect claims,
  • exaggerated claims,
  • unsupported conclusions,
  • outdated information,
  • invented details,
  • misleading wording,
  • incorrect numbers,
  • incorrect dates,
  • incorrect names or titles,
  • fabricated or inaccurate quotes,
  • weak causal claims,
  • claims presented with more certainty than the evidence supports.

The goal is not to criticize the writing style.

The goal is to determine which claims are reliable, which need qualification, and which are false or unverifiable.

Independence Requirement

Whenever possible, perform the review using:

  • a fresh subagent,
  • a new context,
  • or a separate verification pass that does not rely on the original reasoning.

Do not inherit the original answer’s assumptions as facts.

Treat the original work as an unverified source.

Do not optimize for agreement.

Core Mindset

Approach the material with a skeptical but fair mindset.

Assume that at least 1 material claim may be wrong, incomplete, exaggerated, outdated, or unsupported.

Actively try to find it.

This does not mean inventing errors.

Do not flag a correct claim merely to satisfy the assumption.

The final judgment must follow the available evidence.

What to Check

Check all externally verifiable claims, especially:

Numbers

Verify:

  • prices,
  • percentages,
  • statistics,
  • totals,
  • calculations,
  • measurements,
  • dates,
  • durations,
  • rankings,
  • quantities,
  • financial figures.

Check whether:

  • the value is correct,
  • the unit is correct,
  • the time period is clear,
  • the denominator is provided,
  • the number is current,
  • rounding materially changes the meaning.

Names and Identities

Verify:

  • people,
  • companies,
  • products,
  • organizations,
  • locations,
  • job titles,
  • office holders,
  • authors,
  • researchers,
  • laws,
  • standards.

Check spelling, role, attribution, and whether the identity was confused with another entity.

Dates and Timelines

Verify:

  • publication dates,
  • event dates,
  • release dates,
  • deadlines,
  • historical sequences,
  • current availability,
  • chronological relationships.

Distinguish between:

  • when something happened,
  • when it was announced,
  • when an article was published,
  • when a source was last updated.

Quotes and Attributions

Verify:

  • exact wording,
  • speaker,
  • source,
  • context,
  • translation,
  • whether the quote is complete or selectively shortened.

Never treat paraphrased text as a direct quote.

If the original source cannot be located, mark the quote as unverified.

Broad Claims

Closely inspect statements using terms such as:

  • always,
  • never,
  • everyone,
  • nobody,
  • proven,
  • guaranteed,
  • best,
  • worst,
  • only,
  • completely,
  • definitely,
  • universally,
  • scientifically established.

Determine whether the evidence actually supports such certainty.

Causal Claims

Distinguish between:

  • correlation,
  • association,
  • expert opinion,
  • plausible mechanism,
  • demonstrated causation.

Do not accept statements of the form “X causes Y” when the evidence only shows that X and Y are related.

Current Claims

Verify all claims that may have changed, including:

  • prices,
  • laws,
  • regulations,
  • product features,
  • software versions,
  • office holders,
  • schedules,
  • availability,
  • company policies,
  • scientific consensus,
  • market data.

Use current sources when the claim concerns the present.

Evidence Standards

Prefer evidence in this order:

  1. primary sources,
  2. official documentation,
  3. original research,
  4. government or institutional sources,
  5. reputable reporting,
  6. high-quality secondary analysis.

Avoid relying primarily on:

  • unsourced blog posts,
  • search-result snippets,
  • social-media posts,
  • affiliate pages,
  • AI-generated summaries,
  • copied quote collections,
  • websites that do not identify their sources.

A source being widely repeated does not prove that the claim is true.

Source Requirements

For each disputed claim:

  • cite the source that contradicts or qualifies it,
  • cite the source that supports the correction,
  • include the relevant date when freshness matters,
  • separate confirmed evidence from inference.

When no reliable source is available, say so.

Do not fabricate a correction.

Do not replace an unverified claim with another unverified claim.

Verification Statuses

Assign exactly 1 status to each reviewed claim.

❌ False

Reliable evidence directly contradicts the claim.

⚠️ Misleading

The claim contains some truth but omits important context, uses distorted framing, or overstates the evidence.

🟠 Unsupported

The claim may be true, but the available work provides insufficient evidence.

🕒 Outdated

The claim may previously have been correct but is no longer current.

❓ Unverifiable

The claim cannot be reliably confirmed from available evidence or sources.

✅ Supported

Reliable evidence supports the claim as written.

Use ✅ Supported selectively. Focus the report on claims that require correction or qualification.

Severity Levels

Assign a severity to every flagged claim.

Critical

The claim could materially affect:

  • health,
  • safety,
  • legal decisions,
  • financial decisions,
  • security,
  • public reputation,
  • major strategic decisions.

High

The claim substantially changes the conclusion or recommendation.

Medium

The claim is materially inaccurate but does not invalidate the entire work.

Low

The issue is minor, such as a small date discrepancy, imprecise wording, or non-critical attribution problem.

Required Workflow

Step 1: Extract Claims

Identify concrete factual claims from the material.

Separate facts from:

  • opinions,
  • recommendations,
  • predictions,
  • rhetorical language,
  • personal preferences.

Do not fact-check subjective preferences as though they were objective claims.

Step 2: Prioritize Risk

Check the highest-risk claims first:

  1. health and safety,
  2. legal and financial,
  3. security,
  4. numbers and calculations,
  5. quotes and attribution,
  6. current facts,
  7. broad generalizations,
  8. minor factual details.

Step 3: Verify Independently

Search for evidence without copying the original answer’s framing.

Use multiple independent sources for consequential or disputed claims where practical.

Step 4: Try to Disprove the Claim

Look for:

  • contradictory primary evidence,
  • exceptions,
  • changed circumstances,
  • missing qualifications,
  • methodological limitations,
  • alternative interpretations.

Step 5: Classify the Result

Assign:

  • verification status,
  • severity,
  • confidence.

Step 6: Correct or Define the Check

For every flagged claim, provide either:

  • an accurate corrected version,
  • a safer qualified version,
  • or the exact evidence required to verify it.

Step 7: Assess the Whole Work

State whether the overall work is:

  • reliable,
  • mostly reliable with corrections,
  • materially unreliable,
  • or impossible to verify.

Confidence Levels

Use 1 of these confidence levels:

  • High: Strong primary or authoritative evidence exists.
  • Medium: Good evidence exists, but some uncertainty remains.
  • Low: Evidence is limited, conflicting, indirect, or incomplete.

Confidence describes confidence in the fact-check result, not confidence in the original claim.

Required Output Format

Fact-Check Report

Overall Assessment

State the overall reliability in 2–4 sentences.

Include:

  • how many material claims were checked,
  • how many were flagged,
  • whether the main conclusion remains reliable.

Findings

For each flagged claim, use:

Finding <number> — <short title>

Original claim:

<exact claim or faithful short paraphrase>

Status: <verification status> Severity: <Critical | High | Medium | Low> Confidence: <High | Medium | Low>

Why it is questionable: Explain the issue clearly and specifically.

Evidence: Summarize what reliable sources establish.

Accurate version: Provide a corrected or properly qualified version.

Required check: When the claim cannot yet be resolved, state exactly what source, dataset, document, date, test, or confirmation is needed.

Confirmed Claims

Optionally list important claims that were checked and supported.

Do not repeat every trivial claim.

Remaining Uncertainty

List unresolved questions, unavailable sources, conflicting evidence, or freshness concerns.

Final Verdict

Use exactly 1:

  • ✅ Reliable
  • 🟡 Mostly reliable with corrections
  • 🟠 Materially unreliable
  • 🔴 Not reliable
  • ❓ Insufficient evidence

Add 1 concise sentence explaining the verdict.

Compact Output Format

For shorter reviews, use:

ClaimStatusProblemCorrect version or required check
<claim><status><issue><correction or check>

Follow the table with an overall verdict.

Rules Against False Positives

Do not flag a claim solely because:

  • it lacks a citation when it is common knowledge,
  • it is phrased differently from a source,
  • multiple valid interpretations exist,
  • the claim is a clearly marked opinion,
  • the claim is a clearly marked prediction,
  • the evidence is incomplete but still broadly supportive.

When the wording is too strong but the underlying idea is supported, classify it as misleading rather than false.

Rules Against Hallucinated Corrections

Never:

  • invent a source,
  • invent a quote,
  • invent a date,
  • invent a corrected number,
  • claim to have checked a source that was not accessed,
  • use a search snippet as conclusive evidence,
  • present inference as confirmed fact,
  • silently substitute a different claim.

When evidence is unavailable, use ❓ Unverifiable or 🟠 Unsupported.

High-Stakes Claims

For medical, legal, financial, security, or safety claims:

  • prioritize official and primary sources,
  • verify that the source applies to the correct jurisdiction,
  • verify that the source is current,
  • avoid definitive advice based on incomplete evidence,
  • clearly state important limitations.

Do not treat a disclaimer as a substitute for verification.

Handling Quotes

For every direct quote, confirm:

  1. the exact words,
  2. the correct speaker,
  3. the original source,
  4. the publication or event date,
  5. the surrounding context.

When exact wording cannot be confirmed:

The attribution may be plausible, but the exact quote could not be verified from a primary source.

Handling Statistics

For every material statistic, check:

  • original dataset,
  • sample size,
  • collection date,
  • population,
  • geographic scope,
  • methodology,
  • absolute versus relative values,
  • whether the figure is being compared fairly.

A correct number used in the wrong context must be marked misleading.

Handling Missing Sources

When the original work provides no sources:

  • do not assume the claims are false,
  • independently search for support,
  • mark unresolved claims as unsupported or unverifiable,
  • explain what source would be sufficient.

Handling Conflicting Sources

When reliable sources disagree:

  • present both positions,
  • describe the source quality,
  • identify the reason for the conflict when possible,
  • avoid selecting a side without sufficient evidence,
  • reduce confidence accordingly.

Base Prompt

Fact-check the work above with a skeptical, evidence-first approach.

Your task is to identify claims that are false, misleading, unsupported, outdated, unverifiable, misquoted, exaggerated, or fabricated.

Independently verify numbers, names, dates, quotes, attributions, causal claims, broad generalizations, and any facts that may have changed.

Do not agree with the original work by default, and do not inherit its assumptions. Treat it as an unverified source.

Assume that at least 1 material claim may be wrong and actively try to find it, but do not invent errors merely to satisfy that assumption.

For every flagged claim:

1. quote or faithfully identify the original claim,
2. assign a verification status,
3. assign a severity and confidence level,
4. explain exactly why it is questionable,
5. summarize the available evidence,
6. provide an accurate or properly qualified version,
7. or state the exact source, dataset, document, or test needed to verify it.

Do not fabricate sources or corrections. If reliable evidence is unavailable, mark the claim as unsupported or unverifiable.

Finish with an overall reliability verdict.

Strict Prompt

Act as an independent adversarial fact checker.

Review the material above as though it may contain confident hallucinations.

Hunt specifically for:

- incorrect numbers,
- false names or titles,
- wrong dates,
- invented quotes,
- outdated claims,
- unsupported superlatives,
- correlation presented as causation,
- exceptions hidden by words such as “always,” “never,” “only,” or “proven,”
- conclusions that are stronger than the underlying evidence.

Use primary and authoritative sources wherever possible.

For each issue, provide:

- original claim,
- status,
- severity,
- confidence,
- explanation,
- evidence,
- corrected wording or exact verification requirement.

Do not reward plausibility. Require evidence.

Do not invent a correction when the truth cannot be established.

Conclude with exactly 1 verdict:

✅ Reliable  
🟡 Mostly reliable with corrections  
🟠 Materially unreliable  
🔴 Not reliable  
❓ Insufficient evidence

Final Self-Check

Before completing the fact-check, verify:

  • Did I independently test the original assumptions?
  • Did I check the most consequential claims first?
  • Did I distinguish false from merely unsupported?
  • Did I verify numbers, names, dates, and quotes?
  • Did I check whether current claims are still current?
  • Did I distinguish correlation from causation?
  • Did I use primary or authoritative sources where possible?
  • Did I clearly label inference and uncertainty?
  • Did I avoid inventing corrections?
  • Did I provide an actionable verification step for unresolved claims?
  • Does the final verdict match the findings?

Shared output profile

Compose the research profile from output-templates. Keep each claim, source, status, correction, confidence, and remaining uncertainty traceable. Preserve stricter domain rules above, omit empty components, and never fill a component with invented evidence.

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.