AI SEARCH (GEO)

Pair every claim with a source

Use before publishing when a draft is full of numbers you have not checked.

claim-source-pairing.md

Download .md

You are a fact-checking editor preparing a draft for publication.

Draft: {{DRAFT_TEXT}}
Sources I already have: {{AVAILABLE_SOURCES}}
Topic: {{TOPIC}}

Produce a table with one row per factual claim in the draft:
| # | Claim (verbatim) | Type (number / date / attribution / general fact) | Supported by which source | Verdict (supported / partially supported / unsupported / unverifiable) | Suggested fix |

Then:
- List every unsupported number in the draft as a separate block headed REMOVE OR SOURCE.
- Suggest, for each unsupported claim, either a softer phrasing that is defensible without a source, or the type of source that would settle it.

Constraints:
- Do not supply statistics from your own memory to fill gaps. If you believe a figure exists, describe the kind of source rather than stating a value.
- Treat vague quantifiers ("most", "many", "the majority") as claims and flag them.
- Attributions to unnamed groups such as "experts say" are unsupported by default.
- Verdicts must be one of the four words given. No hedged verdicts.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{DRAFT_TEXT}}
{{AVAILABLE_SOURCES}}
{{TOPIC}}

Getting a better result

Unsourced numbers are the main reason an assistant will paraphrase rather than quote you.
Link the source inline next to the claim, not in a references list at the bottom.
Rerun on the edited draft; fixes often introduce new soft quantifiers.

Questions about this prompt

When in the process should I run this?

After the draft is finished and before the edit, on any piece containing numbers somebody else wrote. It is not a research step and it will not find sources for you. It separates the claims you can defend today from those needing either a citation or more careful wording before publication.

What do I need in front of me?

The full draft and the sources you already hold, listed properly rather than as a vague assurance that the writer checked. The prompt is barred from supplying figures from memory, so an empty {{AVAILABLE_SOURCES}} returns a page of unsupported verdicts, which is accurate but tells you nothing new.

What comes back, and which part is the gate?

A row per factual claim with one of four verdicts and a suggested fix, then a REMOVE OR SOURCE block listing every unsupported number. That block is the publication gate. Everything in it is either sourced before the page goes live or rewritten as a claim you can defend without one.

What is the mistake that costs me?

Treating the flagged quantifiers as pedantry. Most, many and the majority are claims, so swapping an unsourced number for one of them moves the problem rather than solving it. Rerun on the edited draft, because the softening pass is precisely where new vague quantifiers get introduced.