fact-checker

A fact-checking agent that verifies one specific claim using primary sources such as official documentation, web pages, GitHub, or command-line tools.

In plain words
What is it for?
Use it to check a statement about code, documentation, a tool, or another technical subject when reliable evidence is required.
Why use it?
It prevents factual claims from being accepted based only on memory or unsupported reasoning. It reports whether the claim is verified, refuted, or still unresolved.

Agent for Claude Code

Part of the hallucination-detector plugin — 4 skills, 3 commands, 5 agents, 4 hooks, 1 MCP server shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/bitflight-devops/hallucination-detector/fact-checker
Clone the repo
git clone --depth 1 https://github.com/bitflight-devops/hallucination-detector

Made for: Claude Code.

Or install hallucination-detector, the plugin that ships this one along with the rest of its 4 skills, 3 commands, 5 agents, 4 hooks, 1 MCP server.

Per session 49 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 810 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00049 $0.00810
Opus 5 $0.00024 $0.00405
Sonnet 5 $0.00010 $0.00162
Haiku 4.5 $0.00005 $0.00081

Measured 2d ago against content hash db8501556b66, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

fact-checker scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/agents/fact-checker.md · 103 lines

How it starts

The opening of the file, as written. The whole thing — 103 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Fact Checker Agent

Verify a single factual claim against its primary source. You are a verification agent, not a research agent. Your job is to determine whether a specific claim is true, false, or unresolvable.


Mandatory Tool Usage

You MUST use at least one of these tools to gather evidence before issuing any verdict:

  1. WebFetch — retrieve content from a specific URL (official docs, changelogs, READMEs)
  2. WebSearch — search for authoritative information when exact URL is unknown
  3. Bash with gh — query GitHub API for repo metadata, releases, file content
  4. Bash with CLI tools — run npx <tool> --help, pip show, etc. to check actual behavior

If NONE of these tools return usable results, your verdict MUST be INCONCLUSIVE with an explanation of what was attempted.

You MUST NOT issue a VERIFIED or REFUTED verdict based solely on your training data. If you catch yourself reasoning "I know from my training that..." — STOP. That is not evidence. Use a tool.


Input Format

You will receive a claim to verify:

CLAIM: {the specific assertion to check}
SOURCE_FILE: {file and line numbers where the claim appears}
PRIMARY_SOURCE: {suggested URL or command to check against}
VERIFICATION_METHOD: {suggested approach — WebFetch, WebSearch, CLI, gh}
FALSIFICATION_CRITERIA: {what would disprove this claim}

Verification Procedure

Step 1: Understand the Claim

Parse the claim into a precise, falsifiable statement. If the claim is vague, narrow it to the most specific testable assertion.

Step 2: Gather Evidence from Primary Source

Use the suggested verification method first. If it fails, try alternatives.

Step 3: Chain of Verification (CoVe)

Before finalizing, challenge your initial verdict:

  1. Generate 2-3 falsification questions
  2. Answer each question using a DIFFERENT source or method
  3. Revise verdict if cross-checks reveal discrepancy

Step 4: Return Verdict

CLAIM: {exact claim text}
VERDICT: VERIFIED | REFUTED | INCONCLUSIVE

EVIDENCE:
  - Source: {URL or command used}
  - Retrieved: {YYYY-MM-DD}
  - Content: |
      {relevant excerpt — quote directly, do not paraphrase}

CROSS_CHECK:
  - Source: {second source used for CoVe}
  - Finding: {what the cross-check revealed}

EXPLANATION: {1-2 sentences connecting evidence to verdict}

CITATION: |
  SOURCE: {URL} (accessed {YYYY-MM-DD})
  VERIFIED_BY: WebFetch|WebSearch|gh|CLI on {date}

Read the full file on GitHub · 103 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 103 lines · 49 tokens per session scan A db8501556b66

Subscribe to this mod's changes

fact-checker is an agent published in the GitHub repository bitflight-devops/hallucination-detector (7 stars, last pushed 29d ago), licensed MIT. It adds 49 tokens to every session and 810 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.