fact-checker
在发布前验证数据内容中的声明。在审查草稿以核实事实准确性、检查代码示例是否正确、验证统计数据和来源、或验证引言和归属时使用。生成包含每个声明裁决的结构化事实核查报告。对于需要扩展研究的深度验证,委托给事实核查子代理。
📖 中文导读:在发布前验证数据内容中的声明。在审查草稿以核实事实准确性、检查代码示例是否正确、验证统计数据和来源、或验证引言和归属时使用。生成包含每个声明裁决的结构化事实核查报告。对于需要扩展研究的深度验证,委托给事实核查子代理。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: fact-checker description: Verify claims in tacosdedatos content before publication. Use this skill when reviewing drafts for factual accuracy, checking code examples work correctly, validating statistics and sources, or verifying quotes and attributions. Produces a structured fact-check report with verdicts for each claim. For deep verification requiring extended research, delegate to the fact-checker subagent instead.
Fact-Checker
Verify technical, statistical, and attribution claims in tacosdedatos content.
References:
references/claim-categories.md— Claim types and verification methodsreferences/report-format.md— Report template and examples
Verification Workflow
Step 1: Extract Claims
Scan the content and extract all verifiable claims:
- Technical claims — Code, APIs, versions, tool capabilities
- Statistical claims — Numbers, percentages, study references
- Attribution claims — Quotes, sources, links
See references/claim-categories.md for identification signals.
Step 2: Prioritize
Focus verification effort on high-impact claims:
| Priority | Examples |
|---|---|
| High | Code readers will copy, statistics supporting arguments, named quotes |
| Medium | Version numbers, tool comparisons |
| Low | Hyperbolic rhetoric, subjective assessments |
Step 3: Verify Each Claim
Technical claims:
1. Code snippets → Execute to verify they run
2. API behavior → Web search for official docs
3. Version claims → Check release notes
Statistical claims:
1. Find original source via web search
2. Verify the number matches
3. Check if data is current
Attribution claims:
1. Web search for original quote/source
2. Verify link validity
3. Confirm attribution accuracy
Step 4: Flag for Human Review
Mark claims that cannot be automatically verified:
- Insider knowledge or personal anecdotes
- Future predictions
- Unpublished/internal data
- Controversial interpretations
Step 5: Generate Report
Produce a structured report using the template in references/report-format.md.
Verdicts:
- VERIFIED — All claims check out
- ISSUES FOUND — Corrections needed before publishing
- NEEDS HUMAN REVIEW — Editor must decide on flagged items
Quick Check vs Deep Verification
| Scope | Use Case | Approach |
|---|---|---|
| Quick check | Pre-publication review | This skill: scan, verify obvious claims, flag concerns |
| Deep verification | Investigative piece, controversial topic | Delegate to fact-checker subagent for extended research |
Code Verification
For code blocks, attempt execution when possible:
# Run code in sandbox
# Capture: success/failure, output, errors
# Report: which blocks run, which fail
Report format for code:
| Code Block | Location | Result |
|---|---|---|
| API example | Section 2 | ✓ Runs |
| Data pipeline | Section 4 | ✗ Error: missing import |
Common Issues in tacosdedatos Content
| Issue | Detection | Recommendation |
|---|---|---|
| Outdated package names | Package not found on pip/npm | Check current package name |
| Deprecated API syntax | Code runs but with warnings | Update to current syntax |
| Broken links | 404 or redirect to unrelated page | Find updated URL or remove |
| Misattributed quotes | Original source says different | Correct attribution or rephrase |
| Stale statistics | Data >2 years old | Find current data or note date |
Output Format
Always produce a report following the template in references/report-format.md.
Minimum report contents:
- Overall verdict (VERIFIED / ISSUES FOUND / NEEDS HUMAN REVIEW)
- Summary of findings
- List of claims checked with individual verdicts
- Code execution results (if applicable)
- Specific recommendations for fixes