Introducing /parse. Convert PDFs, Word docs, or spreadsheets into clean data for AI agents 5x faster. Try it now →

How do I create a fact-checking agent skill?

A fact-checking agent skill can be a callable tool you register with an AI agent that takes a claim as input, searches the web for relevant sources, and returns a verdict (supported, contradicted, or unverified) with citations. The critical part is the search step: an LLM reasoning over its training data alone cannot verify recent or domain-specific claims reliably, because the answer depends on what sources actually say right now, not what the model remembers from pretraining.

ApproachEvidence sourceFreshnessReturns citations
LLM reasoning onlyTraining dataStale, cutoff-limitedNo
RAG over internal docsEmbedded document storeAs fresh as last indexYes, from indexed docs
Web search at query timeLive webCurrentYes, from retrieved pages
Dedicated fact-check APICurated claim databasesVaries by providerPartial

The minimal skill has three steps: (1) turn the claim into a search query, (2) retrieve the top results with full page content, (3) pass the claim and retrieved content to an LLM with a prompt asking it to assess whether the evidence supports or contradicts the claim. The output is a structured verdict with the source URLs used. Multi-source verification — running the same claim against two or three independent results — reduces the chance of a single biased or outdated source skewing the verdict.

Once the Firecrawl CLI is installed, the search step is a single command:

npx -y firecrawl-cli@latest init
firecrawl search "your claim here" --limit 3

The CLI returns full page content per result in markdown, which you pipe directly into your agent's context. A minimal shell-based fact-check skill looks like:

Firecrawl's Search API returns full page content per result, not just snippets, so the LLM has complete article text to reason against rather than 2-3 sentence extracts that often omit the relevant detail. The CLI also installs as a tool inside Claude Code and other agents that support tool use, so agents can invoke firecrawl search directly without any additional wiring.

Last updated: Apr 30, 2026
FOOTER
The easiest way to extract
data from the web
Backed by
Y Combinator
LinkedinGithubYouTube
SOC II · Type 2
AICPA
SOC 2
X (Twitter)
Discord