fact-check

name

fact-check

description

Verify claims with cited live-web sources via the Tavily MCP. Use when the user asks "is this true", "verify this", "fact-check", "find sources for", "is there evidence for", "did X actually say Y", "what are the sources on", or wants to validate something they read or wrote before publishing. Returns a verdict (supported / refuted / mixed / unverifiable) backed by primary sources where possible. For prose research without verification framing, use deep-research.

Fact-Check

Verify a specific claim using live web sources. Output: a verdict, the strongest supporting sources, the strongest refuting sources, and what remains unverifiable.

When to use this skill

• "Is this true: …"
• "Fact-check this paragraph for me."
• "Did X really say Y?"
• "Find sources that support / refute …"
• Pre-publish checks on a draft.

For broad research without verification framing, use deep-research. For breaking-news verification, also pull news-monitor patterns.

Workflow

1. Decompose: break the claim into atomic checkable facts. A sentence usually contains 2–5 of them.
2. For each fact:
   • tavily_search with search_depth: "advanced" and topic: "general" (or news for recent claims).
   • Prefer primary sources: include_domains set to .gov, .edu, official company domains, reputable wires (Reuters, AP), or the source the user named.
   • tavily_extract the top supporting and refuting URLs.
3. Adjudicate each fact:
   • Supported (multiple independent sources agree)
   • Refuted (credible sources contradict)
   • Mixed (sources disagree)
   • Unverifiable (no credible sources found)
4. Synthesize an overall verdict for the claim.

Output format

# Fact-check: "{paraphrased claim}"
*{YYYY-MM-DD}*

## Verdict: {Supported / Refuted / Mixed / Unverifiable}

{1–2 sentence summary of why}

## Atomic facts

### 1. {Fact} — {verdict}
- **Supports:** {brief} [1]
- **Refutes:** {brief} [2]
- **Notes:** {nuance}

### 2. {Fact} — {verdict}
…

## What we couldn't verify
- {Fact} — searched {N} sources, no credible primary source found.

## Sources
[1] {url} — {publisher}, {date}
[2] {url} — {publisher}, {date}

Source quality hierarchy

When sources disagree, weight them:

1. Primary: official statements, regulatory filings, court documents, original research.
2. Authoritative secondary: Reuters, AP, Bloomberg, BBC, peer- reviewed journals.
3. Specialist outlets: trade publications with editorial standards.
4. General media with verification: major newspapers.
5. Aggregators / blogs: only as pointers to primary sources.
6. Social media / forums: noted, but rarely sufficient on their own.

Always tell the user which level of source you're relying on for each verdict.

Things never to do

• Never fabricate citations. If you can't find a source, say so.
• Never silently rely on training knowledge. This skill exists because training data goes stale; if a claim is post-cutoff, you must verify with tavily_* calls.
• Never present a single source as proof. A verdict needs at least two independent sources unless the source is primary (official filing, government data, the company's own release for a fact about themselves).

Quoting rules

When quoting from extracted content, keep quotes short (under 15 words) and always cite. Prefer paraphrase + citation over long block quotes.

Failure modes

• Claim is too vague: ask the user to sharpen it. "Vaccines work" isn't checkable; "{vaccine} reduced {outcome} by {N}% in {study}" is.
• Only finds the original source: an unsupported claim from a single source is not "verified" — say so.
• Time-sensitive claim: include time_range: "month" or narrower so you don't surface stale facts as current.