verify-answer

verify-answer

Decomposes an answer into atomic factual claims and checks each against the evidence the user provides. Returns a clean per-claim report so the user can see exactly what the answer got right, what it got wrong, and what the sources didn't cover.

This skill is prompt-only — no tools, no network, no MCP server required. It works in any Claude installation. Built on the Chain-of-Verification pattern (Dhuliawala et al., 2023).

Looking for the automated version? The full agentic-research-engine-oss pipeline runs this verification step automatically after every synthesized answer, sources included. Install with pip install agentic-research-engine or /plugin marketplace add https://github.com/TheAiSingularity/agentic-research-engine-oss.

---

How to use

Paste an answer and the sources it cites, then invoke:

/verify-answer

Or ask in plain English: "verify this," "fact-check these claims," or "is this supported by the sources?"

What I need from you:

1. The answer — the paragraph, bullet list, or report you want verified.
2. The evidence — the sources it was based on. One of these forms:
   • Copy-pasted text from the sources, OR
   • URLs (I'll note which claims reference which URL but can't fetch them if I don't have browsing), OR
   • A structured list like [1] <source text>, [2] <source text>.

If you only paste the answer without evidence, I'll ask what sources you want it checked against.

---

How I verify

For every answer, I run this protocol:

1. Decompose — I extract each standalone factual claim as one line. A claim is something like "Paxlovid reduced hospitalization by 44 %" — a specific, checkable statement. I ignore editorial hedging, conclusions, and meta-sentences.
2. Check — For each claim, I look through the evidence and mark it:
   • VERIFIED — the claim is directly supported by the evidence.
   • UNVERIFIED — the evidence doesn't mention this claim (not wrong, just unsupported).
   • CONTRADICTED — the evidence explicitly says something different from the claim.
3. Report — Each claim on its own line with its verdict and a one-sentence justification referencing the evidence number.
4. Summary — X/Y claims verified + the ratio + any red flags.