The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
Do a fact-check pass on the draft below. You will not verify anything, because you cannot. Your job is to find every claim a human has to check and lay it out so the checking is fast. ## 1. The draft [THE DRAFT] ## 2. The raw material it was built from [RAW MATERIAL] ## 3. Where it is going [WHERE IT WILL APPEAR] ## 4. Extract every checkable claim Go sentence by sentence. A checkable claim is anything a careful reader could ask "how do you know?" about: a number, a date, a name, a quote, a cause, a comparison, "most", "always", "first", "only", anything about a person or an organization, anything stated as a fact about the world rather than as my opinion or my experience. Put every one in a table with these columns: - Claim: quoted exactly as it appears in the draft. - Type: number, date, name, quote, causal, comparative, generalization, third party, or other. - Where it came from: the line in the raw material that supports it, quoted. If nothing supports it, write "unsourced". If the raw material says something slightly different, write "drifted from:" and quote the original. - If wrong: one line on the consequence. Embarrassing, misleading, harmful to someone, or legally risky. - What would verify it: the specific source or check a human would need. Not "research", but "the March export", "the original survey question", "ask the person quoted". ## 5. Rules - Do not mark anything as verified, confirmed, true, false, likely, or plausible. You do not know. The columns above are the entire output. - Do not fix, soften, or rewrite any claim. If one should probably be cut, say so under "If wrong" and nothing more. - Include claims that came from my own notes. My notes can be wrong too. - Include my opinions only where they are phrased as facts about the world. - Flag every quote whose wording differs from the raw material, even by one word. - Do not skip a claim because it seems obvious. Obvious is how wrong things get published. ## 6. After the table - The three claims where "unsourced" meets the worst "if wrong", in order of priority. - Then this sentence, exactly: "I have not verified any of these. This is a list for you to check." - Then stop.
- Best for
- Any draft going somewhere public or consequential: a post, a newsletter, a report, an email to a client
- Drafts that mix your own notes with the assistant's drafting, where you have lost track of which numbers were yours
- Quotes. The table flags any wording that drifted from the raw material, even by one word
- The step between critique and final edit, when the piece is stable enough that the claims will not change again
- Not for
- Actually verifying anything. The table tells you what to check and where. You open the sources
- Pieces with no factual claims at all. Run it anyway if you are not sure: 'most people' is a claim
- Replacing a professional fact-checker, legal review, or expert read where the stakes call for one
An assistant cannot check facts. It can say 'this appears accurate' with total composure about a figure it generated a minute earlier. The fact-check lesson on this site explains why, and what AI-assisted drafting does to the ordinary habit of trusting a sentence because it reads well.
What an assistant can do, and do well, is read a draft sentence by sentence and pull out every claim a careful reader could ask 'how do you know?' about. Numbers, dates, names, quotes, causes, comparisons, the word 'most'. It can then match each one against the raw material you gave it and say which claims came from your notes, which drifted from them, and which came from nowhere. That is a list, not a verdict, and it turns an afternoon of rereading into an hour of checking.
This prompt produces that list as a table, with two extra columns that make the checking faster: what happens if the claim is wrong, and what kind of source would settle it. It ends with the three worst unsourced claims and one fixed sentence: 'I have not verified any of these. This is a list for you to check.' If the assistant says anything else about truth, the prompt has failed.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[THE DRAFT]RequiredThe draft as it stands, pasted whole, including headline and subheads.
ExampleThe 900-word post about our clinic's no-show rate, pasted in full
[RAW MATERIAL]RequiredEverything the draft was built from: notes, exports, documents, quotes as you recorded them. The table matches claims against this.
ExampleThe scheduling system export showing no-shows at 31 percent in March and 12 percent in August; the text-reminder vendor's invoice; my front desk coordinator's words as I wrote them down: 'a lot of people just forget'; the date we switched (May 4)
[WHERE IT WILL APPEAR]The venue and audience, which set what 'if wrong' costs.
ExampleMy clinic's blog and a newsletter to about 1,200 patients; other clinic owners read it too
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
'You will not verify anything, because you cannot' in the first line
Names the limit before the assistant can perform past it. Framing the job as list-making rather than judging is what keeps 'this appears accurate' out of the output.
Block 2: the raw material
The comparison set. Without it, the 'where it came from' column is impossible and the assistant will guess at provenance, which is worse than no column at all.
Block 4: the definition of a checkable claim
The list of trigger words ('most', 'first', 'only') catches the claims that hide inside adjectives. 'Stated as a fact about the world rather than as my opinion' catches the rest, including the ones you wrote yourself.
The five columns
Provenance is the key column: 'unsourced' and 'drifted from' are the two findings that matter most. 'If wrong' ranks the rows. 'What would verify it' turns each row into a task with a named source instead of a vague worry.
Block 5: the forbidden verdict words
Listing 'likely' and 'plausible' alongside 'verified' matters, because 'plausible' is the assistant's favorite way to approve something without claiming to. Every one of those words is a judgment it cannot make.
Block 6: the priority three and the fixed sentence
Tells you where to start, and leaves a line in the chat that says, in writing, what the assistant did not do. If you share the chat with a colleague, that sentence is the disclosure.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
A physical therapist's 900-word post about cutting no-shows at her clinic with text reminders, checked against her scheduling export, the vendor invoice, a colleague's words as she wrote them down, and the switch date. Going to the clinic blog and a 1,200-patient newsletter.
| Claim | Type | Where it came from | If wrong | What would verify it | |---|---|---|---|---| | "our no-show rate fell from 31 percent in March to 12 percent in August" | number | "export: no-shows 31% March, 12% August" | Misleading to other clinic owners who may copy the change; embarrassing if the export was filtered | The export itself with its filter settings; confirm both months use the same definition of no-show | | "Studies show reminder texts cut no-shows roughly in half" | generalization, third party | unsourced | Attributes a finding to unnamed research; readers will repeat it | A specific study, read in full, with its population and effect size; or cut the sentence | | "As our front desk coordinator put it, 'most people just forget'" | quote | drifted from: "a lot of people just forget" | Misquotes a named colleague in front of 1,200 patients; "most" is a stronger claim than she made | Ask her; use her words | | "We switched on May 4" | date | "date we switched (May 4)" | Minor if off by days; matters if the drop began before the switch | The vendor activation date on the invoice | | "The reminders cost less than one missed appointment per week" | comparative, causal | unsourced (the invoice amount is in the notes; the value of a missed appointment is not) | Misleading; this is the core cost argument of the piece | The invoice total divided by weeks, against the average billed value of one appointment from the practice system | | "Patients told us they appreciated the reminders" | third party | unsourced | Mild; implies feedback that may not exist | Any written feedback; otherwise rephrase by hand to what you actually heard | Highest priority, unsourced and costly if wrong: 1. "Studies show reminder texts cut no-shows roughly in half" 2. "The reminders cost less than one missed appointment per week" 3. "Patients told us they appreciated the reminders" I have not verified any of these. This is a list for you to check.
Six rows from a 900-word draft, which is typical. The quote row is the one the writer would have missed: 'most' versus 'a lot of' is one word, and it turns a colleague's observation into a statistic. The 'studies show' sentence was the assistant's own addition from the drafting session, and the table exposed it as unsourced without anyone needing to remember who wrote what. She cut it. The cost comparison turned out to be true, and she now had the arithmetic to put in the piece instead of the assertion.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
When you have checked some rows yourself
Sources for rows 1 and 4: [PASTE OR DESCRIBE]. Update only the 'where it came from' column for those rows. Do not mark them verified; I did that, not you.
When a quote needs restoring to the exact words
Row 2 is cut. Row 3: her exact words were 'a lot of people just forget'. Show OLD and NEW for that one sentence, nothing else.
When you want the claims the draft implies but does not state
Now list the implied claims: things a reader would take as true from the draft that it never states outright. Same columns, same rules.
When the body is done and the furniture is not
Same pass on the headline, subheads, and image captions only.
When the piece will face a hostile reader
Which rows would a hostile reader go after first? Reorder the priority list for that reader, with one line on why for each.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
The assistant says a claim 'appears accurate' or 'is consistent with industry data'.
- Cause
- The verdict habit. It is answering the question you did not ask.
- Fix
- Reply: 'Section 5. Remove every judgment of truth. Replace each with unsourced or the raw material line.' If it keeps happening, start a fresh session.
Rows are missing because claims were phrased as opinion ('I think reminders are the cheapest fix there is').
- Cause
- The opinion exemption was read too broadly.
- Fix
- Reply: 'Include any opinion that contains a comparison, a number, or a claim about other people.'
'Where it came from' cites a raw material line that only mentions the topic.
- Cause
- Loose matching. The assistant found the neighborhood and called it the address.
- Fix
- Reply: 'The cited line must contain the same number, name, or wording. Otherwise write drifted from, or unsourced.'
Too few rows: the table summarizes ('the statistics in paragraph 2').
- Cause
- Compression. A table with twelve rows looks like a lot to produce.
- Fix
- Reply: 'One row per claim. Paragraph 2 alone has at least four.' Then count them yourself.
The 'If wrong' column proposes new wording.
- Cause
- The urge to fix.
- Fix
- Reply: 'Describe the consequence. Do not propose wording. If a claim should go, say cut.'
Before you publish
The prompt cannot do these. You can.
- You do the checking. Open the export, the invoice, the study. The table is a map, not a certificate.
- 'Sourced' means a line in your notes exists, not that the line is right. Check your own numbers with the same care as the assistant's.
- Fix every drifted quote to the exact words, or paraphrase it openly. Never leave a quote one word off.
- The priority list is where to start, not where to stop. Work the whole table before release.
- Anything still unsourced after checking is cut, or marked as your own view, by hand.
- If the stakes are legal, medical, financial, or about a real person, the table is the input to a professional review, not a substitute for one.
Questions
Can the assistant verify any of this if it has web access?
A search result is another thing to check, not a verification. The assistant can fetch a page, but you decide whether the page actually supports the claim, whether the page is reliable, and whether the number means what the draft says it means. Keep the 'not verified' line even when it has searched.
What counts as sourced if my raw material is my own memory?
Write 'from memory' next to that item in the notes, and the table will cite it as such. Then decide, row by row, whether memory is enough for that claim's 'if wrong'. For a date, often yes. For a number other people will repeat, usually no.
How long does the real checking take after this?
It depends on the rows. A 1,000-word piece usually produces five to twelve. Most take a minute each: open the file, confirm the figure. Two or three take a phone call or a search for the original source. The table's point is that you know which is which before you start.
Should I run this before or after the critique?
After the critique and before the final human edit, once the claims have settled. Run it earlier and you will run it again, because editing changes claims. If the critique flagged sentences where the reader stopped believing, those go in first.