AgentSkillsCN

fact-check

对照权威来源核实文档中的事实性陈述,有效识别虚假信息、杜撰语句以及错误归属的言论。

SKILL.md
--- frontmatter
name: fact-check
category: instruction
description: Verify factual claims in documents against authoritative sources. Catches hallucinations, fabricated quotes, and misattributed claims.
allowed-tools: Read,Grep,Glob,WebSearch,WebFetch,Write,TodoWrite
version: 1.0.0
permalink: skills-fact-check
tags:
  - verification
  - research-integrity
  - hallucination-detection

Fact-Check Skill

Purpose: Verify factual claims in documents against authoritative sources. Assume any quotes or references are hallucinated unless demonstrably proven.

When to invoke:

  • User asks to "triple check", "verify", "fact-check" a document
  • Before submitting reviews, reports, or assessments with factual claims
  • When reviewing AI-generated content that cites sources

Core Principle: Burden of Proof

Every factual claim carries a high burden of proof that must be discharged with evidence. Unlike normal reading where we assume good faith, verification requires demonstrating accuracy—not just absence of obvious error.

Why? LLM-generated content confidently produces plausible-sounding details that don't exist: invented statistics, misattributed quotes, fabricated publication dates, non-existent researchers.

Workflow

Phase 1: Identify Claims

Read the document and extract ALL factual claims requiring verification:

Claim TypeExamples
NamesResearchers, institutions, organizations
NumbersSample sizes, percentages, dates, funding amounts
PublicationsPaper titles, journals, publication years
QuotesDirect quotes attributed to sources
CredentialsDegrees, positions, affiliations
EventsPresentations, grants, collaborations
TimelinesDuration claims ("10-year collaboration")

Use TodoWrite to track each claim category.

Phase 2: Identify Authoritative Sources

For each claim, determine what would constitute authoritative verification:

Claim TypeAuthoritative Sources
Researcher detailsUniversity profiles, Google Scholar, dblp, ORCID
PublicationsPublisher websites, DOI links, preprint servers
InstitutionsOfficial websites, LinkedIn (for existence)
Project detailsGrant databases, project websites, research plans
StatisticsPrimary source documents, methodology sections
QuotesOriginal source (book, paper, interview)

Critical: If the document references a primary source (e.g., "research plan PDF"), READ THAT FIRST. It's the authoritative source for claims about the project.

Phase 3: Cross-Reference Each Claim

For each claim:

  1. Search authoritative sources using WebSearch, WebFetch, or Read
  2. Quote exact evidence found (or note absence)
  3. Compare claim against evidence
  4. Classify the result

Phase 4: Classify Results

StatusMeaningFormat
VerifiedClaim matches authoritative sourceCite source with link/page
⚠️ Clarification neededSource exists but details differNote discrepancy
🔍 UnverifiableNo authoritative source accessibleNote what was searched
📝 Professional judgmentOpinion/assessment, not factual claimNote this is not a verification target

Phase 5: Compile Report

Create verification report using template:

markdown
## Verified Claims (Accurate)

| Claim   | Source                  |
| ------- | ----------------------- |
| [claim] | [source with link/page] |

## Claims Requiring Clarification

| Claim   | Issue         | Evidence         |
| ------- | ------------- | ---------------- |
| [claim] | [discrepancy] | [what was found] |

## Unverifiable Claims

| Claim   | Search Attempted  |
| ------- | ----------------- |
| [claim] | [sources checked] |

## Professional Judgments (Not Verifiable)

- [assessment 1]
- [assessment 2]

## Conclusion

[Summary: hallucinations found? / clean? / caveats?]

Phase 6: Save Companion File

Save report as {document-name}-verification.md in same directory as source document.

Common Hallucination Patterns

Watch especially for:

  1. Plausible-sounding statistics - "n=825" sounds specific, verify it
  2. Timeline inflation - "10-year collaboration" - check earliest joint publication
  3. Credential enhancement - Verify actual titles/positions
  4. Quote fabrication - Direct quotes are often invented
  5. Publication conflation - Mixing details from different papers
  6. Institutional misattribution - Wrong university/department

What This Skill Does NOT Do

  • Evaluate quality of claims (that's reviewer judgment)
  • Check methodology soundness (domain expertise)
  • Assess writing quality (editorial review)
  • Verify opinions (professional judgment isn't factual)

This skill verifies factual accuracy only.

Example Invocation

code
User: "Triple check everything in my SNSF review - assume any quotes or references are hallucinated unless proven otherwise."

Agent:
1. Reads review document
2. TodoWrite: Lists all factual claims by category
3. Reads primary source (research plan PDF)
4. WebSearch: Verifies researcher profiles, publications, institutions
5. Cross-references each claim
6. Compiles verification report
7. Saves as {review}-verification.md