Fact Checker

Fact Checker

Verify factual claims in documents and propose corrections backed by authoritative sources.

When to use

Trigger when users request:

• "Fact-check this document"
• "Verify these AI model specifications"
• "Check if this information is still accurate"
• "Update outdated data in this file"
• "Validate the claims in this section"

Workflow

Copy this checklist to track progress:

Fact-checking Progress:
- [ ] Step 1: Identify factual claims
- [ ] Step 2: Search authoritative sources
- [ ] Step 3: Compare claims against sources
- [ ] Step 4: Generate correction report
- [ ] Step 5: Apply corrections with user approval

Step 1: Identify factual claims

Scan the document for verifiable statements:

Target claim types:

• Technical specifications (context windows, pricing, features)
• Version numbers and release dates
• Statistical data and metrics
• API capabilities and limitations
• Benchmark scores and performance data

Skip subjective content:

• Opinions and recommendations
• Explanatory prose
• Tutorial instructions
• Architectural discussions

Step 2: Search authoritative sources

For each claim, search official sources:

AI models:

• Official announcement pages (anthropic.com/news, openai.com/index, blog.google)
• API documentation (platform.claude.com/docs, platform.openai.com/docs)
• Developer guides and release notes

Technical libraries:

• Official documentation sites
• GitHub repositories (releases, README)
• Package registries (npm, PyPI, crates.io)

General claims:

• Academic papers and research
• Government statistics
• Industry standards bodies

Search strategy:

• Use model names + specification (e.g., "Claude Opus 4.5 context window")
• Include current year for recent information
• Verify from multiple sources when possible

Step 3: Compare claims against sources

Create a comparison table:

Status codes:

• ✅ Accurate - claim matches sources
• ❌ Incorrect - claim contradicts sources
• ⚠️ Outdated - claim was true but superseded
• ❓ Unverifiable - no authoritative source found

Step 4: Generate correction report

Present findings in structured format:

## Fact-Check Report

### Summary
- Total claims checked: X
- Accurate: Y
- Issues found: Z

### Issues Requiring Correction

#### Issue 1: Outdated AI Model Reference
**Location:** Line 77-80 in docs/file.md
**Current claim:** "Claude 3.5 Sonnet: 200K tokens"
**Correction:** "Claude Sonnet 4.5: 200K tokens"
**Source:** https://platform.claude.com/docs/en/build-with-claude/context-windows
**Rationale:** Claude 3.5 Sonnet has been superseded by Claude Sonnet 4.5 (released Sept 2025)

#### Issue 2: Incorrect Context Window
**Location:** Line 79 in docs/file.md
**Current claim:** "GPT-4o: 128K tokens"
**Correction:** "GPT-5.2: 400K tokens"
**Source:** https://openai.com/index/introducing-gpt-5-2/
**Rationale:** 128K was output limit; context window is 400K. Model also updated to GPT-5.2

Step 5: Apply corrections with user approval

Before making changes:

1. Show the correction report to the user 2. Wait for explicit approval: "Should I apply these corrections?" 3. Only proceed after confirmation

When applying corrections:

# Use Edit tool to update document
# Example:
Edit(
    file_path="docs/03-写作规范/AI辅助写书方法论.md",
    old_string="- Claude 3.5 Sonnet: 200K tokens（约 15 万汉字）",
    new_string="- Claude Sonnet 4.5: 200K tokens（约 15 万汉字）"
)

After corrections:

1. Verify all edits were applied successfully 2. Note the correction summary (e.g., "Updated 4 claims in section 2.1") 3. Remind user to commit changes

Search best practices

Query construction

Good queries (specific, current):

• "Claude Opus 4.5 context window 2026"
• "GPT-5.2 official release announcement"
• "Gemini 3 Pro token limit specifications"

Poor queries (vague, generic):

• "Claude context"
• "AI models"
• "Latest version"

Source evaluation

Prefer official sources: 1. Product official pages (highest authority) 2. API documentation 3. Official blog announcements 4. GitHub releases (for open source)

Use with caution:

• Third-party aggregators (llm-stats.com, etc.) - verify against official sources
• Blog posts and articles - cross-reference claims
• Social media - only for announcements, verify elsewhere

Avoid:

• Outdated documentation
• Unofficial wikis without citations
• Speculation and rumors

Handling ambiguity

When sources conflict:

1. Prioritize most recent official documentation 2. Note the discrepancy in the report 3. Present both sources to the user 4. Recommend contacting vendor if critical

When no source found:

1. Mark as ❓ Unverifiable 2. Suggest alternative phrasing: "According to [Source] as of [Date]..." 3. Recommend adding qualification: "approximately", "reported as"

Special considerations

Time-sensitive information

Always include temporal context:

Good corrections:

• "截至 2026 年 1 月" (As of January 2026)
• "Claude Sonnet 4.5 (released September 2025)"

Poor corrections:

• "Latest version" (becomes outdated)
• "Current model" (ambiguous timeframe)

Numerical precision

Match precision to source:

Source says: "approximately 1 million tokens" Write: "1M tokens (approximately)"

Source says: "200,000 token context window" Write: "200K tokens" (exact)

Citation format

Include citations in corrections:

> **注**：具体上下文窗口以模型官方文档为准，本书写作时使用 Claude Sonnet 4.5 为主要工具。

Link to sources when possible.

Examples

Example 1: Technical specification update

User request: "Fact-check the AI model context windows in section 2.1"

Process: 1. Identify claims: Claude 3.5 Sonnet (200K), GPT-4o (128K), Gemini 1.5 Pro (2M) 2. Search official docs for current models 3. Find: Claude Sonnet 4.5, GPT-5.2, Gemini 3 Pro 4. Generate report showing discrepancies 5. Apply corrections after approval

Example 2: Statistical data verification

User request: "Verify the benchmark scores in chapter 5"

Process: 1. Extract numerical claims 2. Search for official benchmark publications 3. Compare reported vs. source values 4. Flag any discrepancies with source links 5. Update with verified figures

Example 3: Version number validation

User request: "Check if these library versions are still current"

Process: 1. List all version numbers mentioned 2. Check package registries (npm, PyPI, etc.) 3. Identify outdated versions 4. Suggest updates with changelog references 5. Update after user confirms

Quality checklist

Before completing fact-check:

• [ ] All factual claims identified and categorized
• [ ] Each claim verified against official sources
• [ ] Sources are authoritative and current
• [ ] Correction report is clear and actionable
• [ ] Temporal context included where relevant
• [ ] User approval obtained before changes
• [ ] All edits verified successful
• [ ] Summary provided to user

Limitations

This skill cannot:

• Verify subjective opinions or judgments
• Access paywalled or restricted sources
• Determine "truth" in disputed claims
• Predict future specifications or features

For such cases:

• Note the limitation in the report
• Suggest qualification language
• Recommend user research or expert consultation

Next Step: Export Verified Content

After fact-checking, suggest exporting the verified document:

Fact-check complete: [N] claims verified, [M] corrections proposed.

Options:
A) Export as PDF — run /pdf-creator (Recommended for formal documents)
B) Create slides — run /ppt-creator from verified content
C) No thanks — I'll use the corrected document directly

Fact Checker

Verify factual claims in documents using web search and official sources, then apply corrections with user confirmation.

Features

• ✅ Comprehensive fact verification across multiple domains
• 🔍 Searches authoritative sources (official docs, API specs, academic papers)
• 📊 Generates detailed correction reports with sources
• 🤖 Auto-applies corrections after user approval
• 🕐 Adds temporal context to prevent information decay

Supported Claim Types

• AI Model Specifications: Context windows, pricing, features, benchmarks
• Technical Documentation: API capabilities, version numbers, library features
• Statistical Data: Metrics, benchmark scores, performance data
• General Facts: Any verifiable factual statement

Usage Examples

Example 1: Update Outdated AI Model Info

User: Fact-check the AI model specifications in section 2.1

What happens: 1. Identifies claims: "Claude 3.5 Sonnet: 200K tokens", "GPT-4o: 128K tokens" 2. Searches official documentation for current models 3. Finds: Claude Sonnet 4.5, GPT-5.2 with updated specs 4. Generates correction report with sources 5. Applies fixes after user confirms

Example 2: Verify Technical Claims

User: Check if these library versions are still current

What happens: 1. Extracts version numbers from document 2. Checks package registries (npm, PyPI, etc.) 3. Identifies outdated versions 4. Suggests updates with changelog references

Example 3: Validate Statistics

User: Verify the benchmark scores in this section

What happens: 1. Identifies numerical claims and metrics 2. Searches official benchmark publications 3. Compares document values vs. source data 4. Flags discrepancies with authoritative links

Workflow

The skill follows a 5-step process:

Fact-checking Progress:
- [ ] Step 1: Identify factual claims
- [ ] Step 2: Search authoritative sources
- [ ] Step 3: Compare claims against sources
- [ ] Step 4: Generate correction report
- [ ] Step 5: Apply corrections with user approval

Source Evaluation

Preferred sources (in order): 1. Official product pages and documentation 2. API documentation and developer guides 3. Official blog announcements 4. GitHub releases (for open source)

Use with caution:

• Third-party aggregators (verify against official sources)
• Blog posts and articles (cross-reference)

Avoid:

• Outdated documentation
• Unofficial wikis without citations
• Speculation and rumors

Real-World Example

Before:

AI 大模型的"上下文窗口"不断升级：
- Claude 3.5 Sonnet: 200K tokens（约 15 万汉字）
- GPT-4o: 128K tokens（约 10 万汉字）
- Gemini 1.5 Pro: 2M tokens（约 150 万汉字）

After fact-checking:

AI 大模型的"上下文窗口"不断升级（截至 2026 年 1 月）：
- Claude Sonnet 4.5: 200K tokens（约 15 万汉字）
- GPT-5.2: 400K tokens（约 30 万汉字）
- Gemini 3 Pro: 1M tokens（约 75 万汉字）

Changes made:

• ✅ Updated Claude 3.5 Sonnet → Claude Sonnet 4.5
• ✅ Corrected GPT-4o (128K) → GPT-5.2 (400K)
• ✅ Fixed Gemini 1.5 Pro (2M) → Gemini 3 Pro (1M)
• ✅ Added temporal marker "截至 2026 年 1 月"

Installation

# Via CCPM (recommended)
ccpm install @daymade-skills/fact-checker

# Manual installation
Download fact-checker.zip and install through Claude Code

Trigger Keywords

The skill activates when you mention:

• "fact-check this document"
• "verify these claims"
• "check if this is accurate"
• "update outdated information"
• "validate the data"

Configuration

No configuration required. The skill works out of the box.

Limitations

Cannot verify:

• Subjective opinions or judgments
• Future predictions or specifications
• Claims requiring paywalled sources
• Disputed facts without authoritative consensus

For such cases, the skill will:

• Note the limitation in the report
• Suggest qualification language
• Recommend user research or expert consultation

Best Practices

For Authors

1. Run regularly: Fact-check documents periodically to catch outdated info 2. Include dates: Add temporal markers like "as of [date]" to claims 3. Cite sources: Keep original source links for future verification 4. Review reports: Always review the correction report before applying changes

For Fact-Checking

1. Be specific: Target specific sections rather than entire books 2. Verify critical claims first: Prioritize high-impact information 3. Cross-reference: For important claims, verify across multiple sources 4. Update regularly: Technical specs change frequently - recheck periodically

Development

Created with skill-creator v1.2.2 following Anthropic's best practices.

Testing:

• Verified on Claude Sonnet 4.5, Opus 4.5, and Haiku 4
• Tested with real-world documentation updates
• Validated correction workflow with user approval gates

Version History

1.0.0 (2026-01-05)

• Initial release
• Support for AI models, technical docs, statistics
• Auto-correction with user approval
• Comprehensive source evaluation framework

License

MIT License - See repository for details

Contributing

Issues and pull requests welcome at daymade/claude-code-skills