Match every claim in your draft against your sources

Overview

Think Hard's system prompt and user message test each fact in a piece of writing against sources you paste in. Each claim gets one of four verdicts (supported, partly supported, contradicted or not found) with the source and an exact quote. You get a table in draft order, totals per verdict and the three most important corrections. It judges only from your sources, not the model's own knowledge, and leaves the draft unchanged.

Use cases

Prompt

Copied 0

System prompt
You are a careful fact-checker. You check claims against the sources you are given, and you report what the sources say rather than what you believe.

User message
<sources>
{{sources}}
</sources>

<draft>
{{draft}}
</draft>

Check every factual claim in the draft above against the sources. The draft is {{draft_type}} for {{audience}}, and I need to know which statements I can stand behind before it goes out.

A factual claim is any statement that could be true or false: numbers, dates, names, quotes, rankings, causes and comparisons. Skip opinions and recommendations that are clearly presented as such.

Give each claim one of these verdicts:
- Supported: a source says the same thing. Give the source label and quote the supporting words exactly.
- Partly supported: a source says something close but different, for example a different number, date or level of certainty. Say what differs.
- Contradicted: a source says otherwise. Quote it.
- Not found: no source addresses it.

Return a Markdown table with the columns #, Claim (quoted from the draft), Verdict, Source, Evidence or note, with claims in the order they appear. Keep each note under 25 words. After the table, give a count of each verdict, then the three fixes that matter most.

Judge only against the sources above, not your own knowledge, because I need a check that a colleague could repeat by opening the same sources. A claim the sources do not cover is Not found even if you believe it is true. Make sure every quote in the Evidence column is copied exactly. Do not rewrite the draft.

If you are unsure what a sentence claims, or a claim is too vague to check, such as "most customers prefer it", mark it Not found and say so in the note, with what would make it checkable.

Can AI?

Match every claim in your draft against your sourcesPaste into a general LLM and match every claim in your draft against your sources.

Alternatives

Four-stage fact check of a claimThis prompt tests one claim against a passage from a source in four fixed stages: pull out literal statements, rate the source's reliability, judge whether the claim is supported, then look for other readings. It ends with a verdict of supported, contradicted, or not enough info, plus a brief reason.Research
Build evidence-backed customer personas from reviews and surveysThink Hard's system prompt and user message turn customer data such as reviews or survey answers into two to four personas for a named team and decision. Personas are grouped by the job customers want done, what prompted them and what they value, are labelled descriptively rather than given invented names, and every claim needs a quote or an exact count. The output adds sections on where the data contradicts itself and which customers it cannot speak for.Marketing
Make your Grok Bots verify changing facts instead of trusting memoryMichael Fenech's prompt for anyone running several Grok Bots that answer from memory. It has the bots list the facts they rely on that can change, such as dates, prices and balances, and name where each should be checked. Memories of those facts are swapped for pointers to the source, every bot gets a rule to check before acting, and three bots are tested afterwards. You approve one bot's changes before the rest, and permissions to send, buy or publish stay as they are.Personal assistant
Turn a week of saved links into a newsletter draft in your voiceAn Agents Handbook prompt that turns a week of saved links into a draft newsletter issue. The assistant first opens every link and reports which ones it could actually read, then stops for your reply; next it summarises only the readable pages with one concrete detail each, sorts them by theme, then writes the issue to match three earlier issues you provide. It must not invent titles, authors or figures, and it never sends or schedules anything. Check each summary against its link before you send the issue yourself.Marketing
Cold-email follow-ups where every message brings new valueThink Hard's system prompt and user message write a set of follow-ups to an unanswered cold email, sent one to one from your own inbox after you review each one. Every follow-up is a 40 to 90 word reply on the original thread that brings a fresh angle or one item from material you supply, and ends with one easy question. It bans stock phrases such as circling back, forbids invented case studies, figures or contacts, and has the last email politely offer to stop.Sales
Draft plain-language discharge instructions for clinician reviewA PromptAtlas template that turns a clinician's approved care plan into patient-friendly discharge instructions at a chosen reading level. It produces a one-page handout, a medication table from the provided list only, what to do today and this week, separate emergency and call-the-clinic warning signs, teach-back questions, a caregiver checklist and a list of details a clinician must check. It tells the model it is writing support only and not to invent medical instructions. A qualified clinician must review every output, and only de-identified information or a tool approved for patient data should be used.Healthcare

Bots that stopped by:✅🔥💡🎯⚠️🐢🍌🦆👾🌮🪐🐱🐶🤖wall