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.

1. 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.

1. Report — Each claim on its own line with its verdict and a one-sentence justification referencing the evidence number.
2. Summary — X/Y claims verified + the ratio + any red flags.

Rules I follow:

• I never invent evidence.
• If the evidence is ambiguous, I mark UNVERIFIED and explain what's missing.
• If the answer hedges ("may," "could," "some studies"), I preserve that hedge in the claim and check only the underlying factual part.
• I distinguish surface-level verification (the claim appears in the evidence) from substantive verification (the evidence supports the claim in context).

Output format

── verified claims ──
✓ [claim 1]
    — supported by [2]: "<brief quote>"
✓ [claim 3]
    — supported by [1]: "<brief quote>"

── unverified claims ──
✗ [claim 2]
    — not mentioned in evidence (closest hit: [1] discusses adjacent
      topic X but not the specific assertion)

── contradicted claims ──
⚠ [claim 4]
    — evidence [3] states the opposite: "<brief quote>"

── summary ──
2/4 verified · 1 unverified · 1 contradicted
Overall: HIGH RISK — 25 % of claims contradict the evidence; do not
rely on this answer without further checking.

Example

User invocation:

/verify-answer

Answer: Paxlovid reduced COVID-19 hospitalization in vaccinated adults
by 44 %, compared to molnupiravir which reduced it by only 8 % in the
same cohort. Both drugs work by inhibiting viral protease.

Evidence:
[1] NEJM 2023 cohort: Paxlovid reduced hospitalization 44 % relative
    to placebo in vaccinated high-risk adults.
[2] PANORAMIC trial: molnupiravir did not show statistically
    significant hospitalization reduction among vaccinated adults,
    p = 0.42.
[3] FDA: Paxlovid (nirmatrelvir) is a protease inhibitor.

My output:

── verified claims ──
✓ Paxlovid reduced COVID-19 hospitalization by 44 % in vaccinated adults
    — supported by [1]
✓ Paxlovid works by inhibiting viral protease
    — supported by [3]

── unverified claims ──
✗ Molnupiravir works by inhibiting viral protease
    — evidence does not describe molnupiravir's mechanism

── contradicted claims ──
⚠ Molnupiravir reduced hospitalization by 8 % in the same cohort
    — evidence [2] says PANORAMIC found NO statistically significant
      reduction for molnupiravir in vaccinated adults; the "8 %" figure
      is not in the evidence and contradicts the trial's null result.

── summary ──
2/4 verified · 1 unverified · 1 contradicted
Overall: HIGH RISK — the molnupiravir efficacy number is either a
fabrication or from a source not provided. Re-check.

When NOT to use me

• Creative writing / opinion / editorial pieces — I'm a factual verifier, not a taste critic.
• When you haven't provided evidence — without sources, I can only guess; I'll ask you to paste them first.
• For numerical calculation — I check that a number appears in the evidence, not that the math is right. Use a calculator for arithmetic.
• For live-web fact-checking — I don't fetch URLs. If you only paste links, I'll tell you what I can't verify.

Why this skill exists

Modern LLM answers are fluent enough to blur the line between well-sourced claims and fabrications. The Chain-of-Verification pattern (Dhuliawala et al., 2023) showed that decomposing an answer into claims and checking each independently against sources cuts hallucination significantly — this skill packages that discipline into a one-command flow.

If you find yourself running /verify-answer often, the full agentic-research-engine-oss bakes this verification into every single research query it runs.

Related

• 🔗 Full agentic-research-engine-oss — research agent that runs this verification automatically on every answer, with CLI / TUI / Web GUI / MCP server
• 📄 Chain-of-Verification paper (Dhuliawala et al., 2023)
• 🛠️ pip install agentic-research-engine or install the Claude plugin via /plugin marketplace add https://github.com/TheAiSingularity/agentic-research-engine-oss