fact-checker

What this skill does

> **Note (roadmap).** Slated to be rebuilt as a dynamic, model-driven workflow rather than a static checklist — keep the intent (a verdict per claim), generalize the mechanics.

# Fact-Checker

Verify technical, statistical, and attribution claims in a draft.

**References**:
- `references/claim-categories.md` — Claim types and verification methods
- `references/report-format.md` — Report template and examples

---

## Verification Workflow

### Step 1: Extract Claims

Scan the content and extract all verifiable claims:

1. **Technical claims** — Code, APIs, versions, tool capabilities
2. **Statistical claims** — Numbers, percentages, study references
3. **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- Unpublished/internal data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 a research subagent for extended verification |

---

## Code Verification

For code blocks, attempt execution when possible:

```python
# 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 to Watch For

| 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 |
| AI-fabricated quotes | Quote sounds plausible but no source confirms wording | Verify against the literal source, not just URL/date |

---

## Output Format

Always produce a report following the template in `references/report-format.md`.

**Minimum report contents:**
1. Overall verdict (VERIFIED / ISSUES FOUND / NEEDS HUMAN REVIEW)
2. Summary of findings
3. List of claims checked with individual verdicts
4. Code execution results (if applicable)
5. Specific recommendations for fixes