Fact-Claim Checker

A public launch post from a small software company, with prices, dates, a user count and a discount, mixed with opinion.

After (skill output, excerpt):
Claims found: 17. High risk: 6. Medium: 7. Low: 4.
"more than 60 people have signed up used Plausible successfully": a number, and "successfully" is undefined. A typo changes the meaning ("signed up used").
"up to 10k pageviews - $6 / mo" versus later "a small (<10k) tier": pick one.
Safer wording: "More than 60 people have signed up for the beta, and [N] of them have [SOURCE: your definition of 'successfully']."

Paste a draft before you publish it. You get a list of the claims that need a source, ranked by risk, with the kind of source that would settle each one, and safer wording for the high-risk ones you cannot back up. It does not tell you what is true. It tells you what to check.

1 real-data run on a public launch post (Claude Sonnet). All 17 flagged claims and both internal conflicts were real.

1 real-data run on a public launch post (Claude Sonnet). We read every flagged claim against the source text: all 17 were real claims, and the two conflicts it raised were real. Not yet run on other models. A second, harder run (2026-10-04, Claude Sonnet, a draft full of unsourced boasts and a request to say which are true): it refused to rule on truth and listed what would settle each claim, but its summary said 8 high-risk claims when the table showed 7. We added a rule and re-ran: the counts now match the table.

Tested on a real example before release. The example is shown in the Before and after section above.

How we test, what we do not do, and which assistants we ran it on.

• It does not look anything up or rule claims true or false.
• It can over-flag low-risk items.
• Not a legal, medical or financial review.

Works with any assistant that reads SKILL.md skills (Claude, Codex, Gemini CLI, Cursor and others), plus a paste-in prompt version for ChatGPT or any chatbot. Pasted text only. No code or web access needed.

1. Run npx skills add proskillpacks/skills in your project and pick this skill. Or copy the fact-claim-checker folder from GitHub (https://github.com/proskillpacks/skills/tree/main/skills/writing/fact-claim-checker).
2. Put the fact-claim-checker folder where your assistant looks for skills. The generic path is .agents/skills/ (or npx skills add from a folder). Tool-specific paths:
   • Claude Code: ~/.claude/skills/ (or .claude/skills/ inside one project)
   • OpenAI Codex, Gemini CLI, Cursor and others: .agents/skills/ or the skills folder their docs name
   • claude.ai: zip the skill folder yourself, then Customize, Skills, the plus button, Create skill, Upload a skill (code execution must be on in Settings, Capabilities)
   • ChatGPT or any chatbot: open the skill's prompt.md, paste it into the chat, then paste your own text or link
   • Steps for each tool, with links to its official docs: how to install Agent Skills
3. Ask for the job in plain words, for example: "Fact-check my draft before I publish it.". The skill loads by itself.