fact-check

Fact-Check Skill

Verify factual claims in written content using 6 parallel verification agents with a tiered source credibility system.

Input Handling

The user provides content in one of three ways:

1. Direct text after /fact-check - use the pasted content as-is
2. File path - read the file with the Read tool
3. URL - fetch the page content with WebFetch

Detect the input type automatically:

• Starts with http:// or https:// → URL, fetch with WebFetch
• Starts with / or ~/ or matches a file path pattern (e.g., path/to/file.md) → file path, read with Read tool
• Otherwise → treat as direct text content

If the input is ambiguous, ask the user to clarify.

WebFetch fallback: If WebFetch fails (paywall, bot protection, timeout), use this fallback chain:

1. Try WebSearch with the URL + key terms from the title to find cached/syndicated copies
2. Search for the article title to find other sources covering the same content
3. If the article can be partially reconstructed from search snippets, proceed with those claims but note in the report: "Article could not be fetched directly. Claims reconstructed from search results and surrounding coverage."
4. If no content can be recovered, tell the user and suggest they paste the text directly.

Workflow

Fast path: If the input contains 3 or fewer claims after extraction, skip agent dispatch (Step 2) and verify all claims directly in the main session using WebSearch. This avoids the overhead of 6 parallel agents for simple content.

Step 1: Extract Claims

Parse the input text and extract two categories:

Explicit claims - factual statements, statistics, quotes, dates, product names, release info Implicit assumptions - unstated background facts the text relies on to make sense

For each claim, note:

• The exact quote from the text
• The claim type (statistic, date, quote, product info, causal, comparison, etc.)

Step 1.5: Classify Risk

Assign a risk level to each claim:

Risk	Criteria	Why it matters
High	Numbers, dates, release info, direct quotes, pricing	Getting these wrong destroys credibility — readers notice and share corrections
Medium	Causality, comparisons, trend descriptions	Subtler errors that mislead without being obviously wrong
Low	General descriptions, subjective opinions, widely known facts	Low consequence if slightly off

Present the extracted claims to the user before proceeding:

Found N claims (X high-risk, Y medium-risk, Z low-risk). Proceeding with verification...

Step 2: Dispatch 6 Verification Agents

Launch all 6 agents in parallel using the Agent tool. Each agent gets:

• The full original text
• The extracted claims list with risk levels
• Any source URLs found in the text

For each agent, read the corresponding prompt file and include it in the agent's instructions:

Agent	Prompt file	Purpose
Source Verify	agents/source-verify.md	Compare claims against their cited sources
Number Check	agents/number-check.md	Verify dates, numbers, proper nouns
Freshness Check	agents/freshness-check.md	Check if time-sensitive info is current
Context Check	agents/context-check.md	Detect distortion, exaggeration, false causation
Link Check	agents/link-check.md	Validate all URLs in the text
Assumption Check	agents/assumption-check.md	Surface and verify unstated assumptions

Read each agent's prompt file from the skill directory before dispatching. The skill directory path is: the directory containing this SKILL.md file.

Important: Use subagent_type: "general-purpose" for each agent. Launch all 6 in a single message to maximize parallelism.

Input scoping: To reduce token cost, scope each agent's input to what it needs:

• Link Check: Only the URLs found in the text, not the full content
• Number Check: Only claims classified as statistics, dates, or proper nouns
• Source Verify: Full text + source URLs
• Others: Full text + all claims

For short content (<1000 words), send full text to all agents.

Agent failure fallback: If any agent fails (timeout, token limit), perform that agent's verification directly in the main session. Never skip a verification category because an agent failed.

Step 3: Cross-Reference Results

After all agents return:

1. Merge findings - combine results from all 6 agents per claim
2. Detect conflicts - if agents disagree on a claim, note the disagreement and include a "Conflicts" section in the report showing which agents disagree and why
3. Check internal consistency - scan for contradictions within the text itself (e.g., "4 options" but only 3 listed, numbers that don't add up, dates that conflict with each other). These are errors that don't require external verification.
4. Assign source tiers - rate the best source found for each claim:

Tier	Type	Examples	Trust Level
Tier 1	Direct source	Official blogs, academic papers, press releases, government docs	High
Tier 2	Trusted media	Major tech/news outlets (Reuters, AP, NYT, Verge, etc.), official reports	Medium-High
Tier 3	Secondary	Personal blogs, community posts, social media	Low
Tier 4	Unverified	No source found, rumors, unofficial leaks	Very Low

5. Auto-escalate low-tier claims - for any claim where only Tier 3-4 sources were found, automatically search for higher-tier sources using WebSearch. If a Tier 1-2 source is found, update the verdict. If not, mark with a warning. Show escalation attempts in the report:

   ### Tier Escalation
   - Claim #3: Tier 3 -> searched for Tier 1-2 -> found [official source] -> upgraded to Tier 1
   - Claim #7: Tier 4 -> searched for Tier 1-2 -> no higher source found -> remains Tier 4 ⚠️
   

Step 4: Generate Verdict

Assign a verdict to each claim:

• Confirmed - verified against Tier 1-2 sources
• Needs Revision - partially correct, or only low-confidence sources available
• Error - contradicted by reliable sources
• Unverifiable - cannot confirm or deny with available sources

Step 5: Output Report

Present the report in this order — summary first, then details. The reader should get the verdict in 10 seconds, then drill into details if they want.

## Fact-Check Report

### Summary
- Claims checked: N (High: A, Medium: B, Low: C)
- Confirmed: X / Error: Y / Needs Revision: Z / Unverifiable: W

### Verdicts

| # | Claim | Risk | Verdict | Source Tier | Details |
|---|-------|------|---------|-------------|---------|
| 1 | "..." | High | Confirmed | Tier 1 | Verified via official blog post |
| 2 | "..." | High | Error | - | Original says X, but official source says Y |
| 3 | "..." | Med  | Needs Revision | Tier 3 | Only blog sources found, no official confirmation |

### Conflicts (if any)
- Claim #5: Source Verify says "match", Freshness Check says "outdated" -> resolved as "Needs Revision" because...

### Tier Escalation (if any)
- Claim #3: Tier 3 -> Tier 1 (found official source)

Keep it concise. Do NOT append per-agent detailed reports after the main report. The agent-level detail is useful for debugging the skill, not for the user. If the user asks for details on a specific claim, provide them on request.

Step 6: Suggest Fixes

For claims marked Error or Needs Revision, provide diff-style corrections:

### Suggested Fixes

1. [Error] Line 12:
   - "OpenAI는 2026년 말까지 AGI를 달성할 것이라고 발표했다"
   + (삭제 권장 — 공식 발표 확인 불가, Tier 4 소스만 존재)

2. [Needs Revision] Line 28:
   - "2025년 봄 출시 예정"
   + "2025년 봄 출시 예정이었으나 2026년 9월로 세 차례 연기됨"
     Source: [공식 블로그 링크]

Key Principles

• Conservative judgment: when in doubt, mark as "Needs Revision" rather than "Confirmed". It is far better to flag something that turns out to be correct than to miss something that is wrong.
• "I don't know" is valuable: "Unverifiable" is a legitimate and useful verdict. The absence of confirmation is information worth reporting.
• Source quality over quantity: one Tier 1 source outweighs ten Tier 3 sources.
• Context matters: a claim can be technically accurate but misleading in context. The context-check agent catches this.