Fact-check list for every claim in a draft

You are a fact-checker at a careful publication. Do not judge whether claims are true from memory. Your job is to build a verification checklist.

Type of text: [e.g. "blog post", "company report", "speech", "social media thread"]
Audience and stakes: [e.g. "public article, legal risk if wrong"]
Sources the author used (if known): [LIST, or "unknown"]

Steps:
1. Extract every checkable claim: numbers, dates, names and titles, quotes, causal statements ("X led to Y"), comparisons ("the largest", "the first") and descriptions of studies or laws.
2. For each claim give, in a table:
   # | Claim (exact words) | Type | Risk if wrong (high/medium/low) | What would verify it (primary source to look for) | Notes
3. Mark as high risk: statistics, quotes, health, legal or financial statements, and anything about a named person or company.
4. Flag vague claims that cannot be checked as written ("experts say", "studies show") and suggest how to make them specific.
5. List claims where the wording goes beyond what a typical source would support (e.g. "proves" instead of "suggests").

If you can browse the web: for high-risk claims only, look for a primary source, give the link and quote the relevant sentence. Mark each as "SUPPORTED", "CONTRADICTED" or "NOT FOUND". If you cannot browse, say so and do not assign these labels.

Draft:
"""
[PASTE TEXT]
"""

Example (excerpt):
| # | Claim | Type | Risk | What would verify it | Notes |
| 1 | "Bike traffic rose 40% in 2025" | statistic | high | City transport department annual count | Check if 40% is for the whole city or one street |
| 2 | "Mayor Lopez called it a turning point" | quote | high | Original interview or press release | Confirm exact words and date |
| 3 | "Studies show cyclists are happier" | vague | medium | — | Name the study or soften to "some surveys suggest" |

Wording note: "proves cycling reduces traffic" — a single year of counts would support "coincided with" at most.

When to use it

Use this prompt before publishing anything that contains facts others will rely on: articles, reports, grant applications, presentations, press releases, newsletters or even a long social media thread. It is especially useful when the draft was written quickly, compiled from several people’s notes, or produced with the help of a writing assistant.

How to customize it

• [TYPE OF TEXT] sets expectations. A speech may use rounded numbers; a report should not.
• [AUDIENCE AND STAKES] decides how strict the check should be. Public, legal or health content deserves the strictest review.
• [SOURCES] helps you go back to the right documents. If the author is you, list what you used while it is still fresh.
• [PASTE TEXT] should be the final or near-final version, since late edits often introduce new errors.

Why it works

The prompt deliberately splits fact-checking into two jobs. Extracting and classifying claims is something language models do well. Judging truth from memory is not: models can be confidently wrong, and their training data has a cut-off date. Asking for a checklist rather than a verdict plays to the strength and avoids the weakness.

Rating risk focuses your limited time on the claims that could do the most harm. The “what would verify it” column points to primary sources, which professional fact-checkers prefer over secondary reports. The browsing step is optional and conditional, because not every assistant can access the web, and those that can still need their quoted sources checked by you. The wording check catches a subtle problem: a true fact described with stronger language than the evidence supports.

Tips and variations

• Keep a simple log of each claim, the source you used and the date you checked it, so corrections later are fast and transparent.
• Work through the high-risk rows first and record the source link for each claim you verify.
• For images and videos in the piece, see our guide to spotting deepfakes.
• Ask the model to rewrite only the sentences you marked as overstated, keeping everything else unchanged.
• Our editorial on why AI coverage needs evidence, not hype explains the standard we apply ourselves.