Check each claim in a draft against the sources you give, quote the passage that backs or contradicts it, and list what to fix, using nothing from outside those sources.
Sample results: runs and scores on this page are illustrations while testing is set up. How we test
Who it is for: Anyone who writes or approves a report, newsletter or briefing and has the source material at hand.
Select a part to see its prompt, input, output and check.
The same job, with and without the template
Sample
One plain prompt
Fact-check this draft against my sources and tell me what is wrong. Use only the sources. (The same draft and sources are pasted below.)
Overall the draft is mostly accurate, with a few small errors.
Membership: the sources show 312 households at the end of the year, not 340.
Loans: the sources show 1,860, not 1,900. This is a minor rounding difference, and "nearly doubled" is fair given the rise from 1,040.
The Elm Street opening on April 2 and the drop in waiting time from 12 days to 5 match the sources.
Volunteer hours: 2,140 is consistent with "more than 2,000".
Survey: 96 of 104 respondents said they would recommend the library, so the claim is broadly accurate.
Largest tool library in the county: this is plausible given the growth described, although the sources do not confirm it.
Suggested corrections: change 340 to 312 and 1,900 to 1,860.
Problem:It calls the survey claim broadly accurate when 8 of 104 said maybe, calls the county claim plausible when the sources say no count was made, and quotes no passage to check.
With this template
The same input, run through the steps.
Result: 7 claims checked: 3 supported, 1 partly supported, 2 contradicted, 1 not in sources.
Fix before you send
Swipe sideways to see all columns
#
Claim (quote)
What the sources say
Suggested wording
1
"Membership grew from 210 households to 340"
210 households at the start of the year, 312 at the end.
Membership grew from 210 households to 312.
2
"loans nearly doubled to 1,900"
Loans: 1,040 the year before, 1,860 this year.
Loans rose from 1,040 to 1,860.
6
"Every member who answered our survey said they would recommend us"
96 of 104 respondents said they would recommend the library; 8 said maybe.
96 of the 104 members who answered our survey said they would recommend us; 8 said maybe.
7
"we are now the largest tool library in the county"
No count of other tool libraries in the county has been made.
Cut it until a source exists.
Holds up
"We opened a second location on Elm Street on April 2"
"the average wait for the ten most-requested tools fell from 12 days to 5"
"Volunteers logged more than 2,000 hours"
Check:Every claim is quoted word for word from the draft: Seven claims quoted word for word from the draft; the opening line about the best year is an opinion and is left out.
Check:Every passage is quoted word for word from the sources: Seven passages, each quoted word for word from the sources.
Check:Each verdict is one of the four allowed words: Seven verdicts, each one of the four allowed words: 3 supported, 1 partly supported, 2 contradicted, 1 not in sources.
Check:Every fix row is filled in: Four rows under Fix before you send, each with all three cells filled.
Check:Every claim in the fix list is in the draft word for word: All four quoted claims are in the draft word for word.
Check:The counts add up to the verdicts: The Result line says 7 claims, 3 supported, 1 partly supported, 2 contradicted and 1 not in sources; the table has four rows, one for each claim that is not supported.
Check:Each suggested wording says only what the sources say: Each suggested wording uses only figures from the sources (210, 312, 1,040, 1,860, 96, 104 and 8).
For your read:You decide what to do with each fix: For the county claim, will you find a count, soften it or cut it?
Runs and scores
Sample
Claude, Sonnet model
October 4, 2026, easy sample
Check:Every claim is quoted word for word from the draft
Three claims quoted word for word from the draft.
Check:Every passage is quoted word for word from the sources
Three passages, each quoted word for word from the sources.
Check:Each verdict is one of the four allowed words
Three verdicts, all supported.
Check:Every fix row is filled in
One row under Fix before you send, with None found in every cell, because every claim is supported.
Check:Every claim in the fix list is in the draft word for word
The fix list quotes no claim (its row says None found), so every quote in it is in the draft.
Check:The counts add up to the verdicts
The Result line says 3 claims checked, all 3 supported; the fix list has no row for a claim.
Check:Each suggested wording says only what the sources say
There is no suggested wording to compare, because every claim is supported.
For your read:You decide what to do with each fix
Is there anything in the three supported claims you would still confirm with someone?
Fixed: Nothing needed fixing: every check passed on the first pass.
Met 8 of 8
1 of these is for your read: the tool hands it to you.
ChatGPT, GPT model
October 5, 2026, tricky sample
Check:Every claim is quoted word for word from the draft
Six claims quoted word for word from the draft.
First pass failed:Every passage is quoted word for word from the sources
The first version gave the passage for the enrollment claim as "Enrollment rose from 400 to 480", which is not in the sources word for word.
Check:Each verdict is one of the four allowed words
Six verdicts, each one of the four allowed words: 4 supported, 1 contradicted, 1 not in sources.
Check:Every fix row is filled in
Two rows under Fix before you send, each with all three cells filled.
Check:Every claim in the fix list is in the draft word for word
Both quoted claims are in the draft word for word.
Check:The counts add up to the verdicts
The Result line says 6 claims checked, 4 supported, 0 partly supported, 1 contradicted and 1 not in sources; the fix list has two rows.
Failed:Each suggested wording says only what the sources say
The suggested wording for the reading claim says "about 6 in 10", a figure the sources do not give (they say 62 of 100).
For your read:You decide what to do with each fix
The district says its results are not final, so will you wait for the October report before using any reading figure?
Fixed: The passages step ran again once with the rule to quote word for word, and it came back with "Enrollment last summer: 400. This summer: 480.", which cleared the first failure. The report step ran again once and the wording still said "about 6 in 10", so the check failed again and the run stopped there and flagged it for you.
Met 7 of 8
Not met:Each suggested wording says only what the sources say: The wording still says about 6 in 10, a figure the sources do not give.
1 of these is for your read: the tool hands it to you.
n8n, import format 2.20
October 6, 2026, typical sample
Check:Every claim is quoted word for word from the draft
Seven claims quoted word for word from the draft; the opening line about the best year is left out as an opinion.
Check:Every passage is quoted word for word from the sources
Seven passages, each quoted word for word from the sources.
First pass failed:Each verdict is one of the four allowed words
Row 2 said "partially supported", and the Verdict column must say supported, partly supported, contradicted or not in sources.
Check:Every fix row is filled in
Four rows under Fix before you send, each with all three cells filled.
Check:Every claim in the fix list is in the draft word for word
All four quoted claims are in the draft word for word.
Check:The counts add up to the verdicts
The Result line says 7 claims checked, 3 supported, 1 partly supported, 2 contradicted and 1 not in sources; the fix list has four rows.
Check:Each suggested wording says only what the sources say
Each suggested wording uses only figures from the sources (210, 312, 1,040, 1,860, 96, 104 and 8).
For your read:You decide what to do with each fix
Will you check the 312 and 1,860 figures against the original report before you send?
Fixed: The verdicts step ran again once with the four allowed words restated, and every row then used one of them.
Met 8 of 8
1 of these is for your read: the tool hands it to you.
The steps, and the check on each
1
Step 1: List every claim in the draft
Pick out each statement a source could prove or disprove, quoted word for word, before any source is read.
Hands back: A numbered table of the claims, each quoted from the draft.
Check:Every claim is quoted word for word from the draft
A plain rule. If it fails, run the step again (the n8n file does this for you).
2
Step 2: Find the passage for each claim
For each claim, quote the words in the sources that bear on it, or say that none do.
Hands back: A table that puts a quoted passage from the sources beside each claim, or None found.
Check:Every passage is quoted word for word from the sources
A plain rule. If it fails, run the step again (the n8n file does this for you).
3
Step 3: Give each claim a verdict
Compare each claim with its passage and give one of four verdicts, with the reason in one sentence.
Hands back: A table that gives each claim one of four verdicts and the reason.
Check:Each verdict is one of the four allowed words
A plain rule. If it fails, run the step again (the n8n file does this for you).
4
Step 4: Write the fact-check report
Turn the verdicts into a count, a list of what to fix with wording taken from the sources, and a list of what holds up.
Hands back: A report with a count line, a table of fixes and a list of the claims that hold up.
Check:Every fix row is filled in
A plain rule. If it fails, run the step again (the n8n file does this for you).
After the last step
Check:Every claim in the fix list is in the draft word for word
A plain rule. If it fails, run the step again (the n8n file does this for you).
Check:The counts add up to the verdicts
A second look by the AI. If it fails, run the step again (the n8n file does this for you).
Check:Each suggested wording says only what the sources say
A second look by the AI. If it fails, run the step again (the n8n file does this for you).
Check:You decide what to do with each fix
You check this one. If it fails, it is flagged for you.
What you need first
The draft to check(needed)
Paste the text as it stands now. Every number, date and name in it will be checked.
The sources to check it against(needed)
Paste the text of the reports, records or notes the draft should rest on. If you have several, paste them one after another with a title above each.
Ways to adapt it
This template checks the draft against your sources, not against the world. A claim marked not in sources may still be true; it only means your sources do not show it.
If you have several documents, give each a title on its own line above its text, so you can find a passage again.
To allow rounding, change the line about partly supported in the verdicts step, for example by saying that a number rounded to the nearest hundred counts as supported.
To add a verdict word, add it to the verdicts step's rules and to its check, both.
For a long draft, run the template on one section at a time so every claim gets a passage.
Change any step or check in the Editor, then download it again.
A short update with seven claims checked against the year's record; two are wrong, one is stretched and one is not in the sources at all.
The draft to check
Our tool library had its best year yet. Membership grew from 210 households to 340, and loans nearly doubled to 1,900. We opened a second location on Elm Street on April 2, and the average wait for the ten most-requested tools fell from 12 days to 5. Volunteers logged more than 2,000 hours. Every member who answered our survey said they would recommend us, and we are now the largest tool library in the county.
The sources to check it against
Annual report data, tool library. Membership: 210 households at the start of the year, 312 at the end. Loans: 1,040 the year before, 1,860 this year. Second location: opened on Elm Street on April 2. Average wait for the ten most-requested tools: 12 days before, 5 days after the second location opened. Volunteer hours logged: 2,140. Member survey: 96 of 104 respondents said they would recommend the library; 8 said maybe. Other libraries: no count of other tool libraries in the county has been made.
The scoring guide
Sample
One criterion for each check, marked met or not met. (The guides in Quality rate a whole piece of work on three levels.) Under each one is how the three sample runs scored.
Every claim is quoted word for word from the draft
A plain rule
Meets:Every claim in the table is found word for word in the draft.
Does not meet:A claim has been reworded or corrected, or is not in the draft word for word.
Claude:Met
ChatGPT:Met
n8n:Met
Every passage is quoted word for word from the sources
A plain rule
Meets:Every passage is found word for word in the sources, or the row says None found.
Does not meet:A passage has been reworded, or is not found word for word in the sources.
Claude:Met
ChatGPT:Met
n8n:Met
Each verdict is one of the four allowed words
A plain rule
Meets:Every claim has exactly one of supported, partly supported, contradicted or not in sources.
Does not meet:A row has another word, such as partially supported or unverified.
Claude:Met
ChatGPT:Met
n8n:Met
Every fix row is filled in
A plain rule
Meets:Every row under Fix before you send has the claim, what the sources say and the suggested wording filled in.
Does not meet:A row under Fix before you send has a blank or a dash in one of its three cells.
Claude:Met
ChatGPT:Met
n8n:Met
Every claim in the fix list is in the draft word for word
A plain rule
Meets:Every claim quoted in the fix list is found word for word in the draft.
Does not meet:A quoted claim has been reworded or is not in the draft.
Claude:Met
ChatGPT:Met
n8n:Met
The counts add up to the verdicts
A second look by the AI
Meets:The numbers on the Result line match the verdicts, and the fix list has one row for each claim that is not supported.
Does not meet:A count on the Result line is wrong, or a claim that is not supported has no row in the fix list, or a supported claim has one.
Claude:Met
ChatGPT:Met
n8n:Met
Each suggested wording says only what the sources say
A second look by the AI
Meets:Every number, date and name in each suggested wording appears in the sources, and the wording says no more than they do.
Does not meet:A suggested wording adds a figure or a fact the sources do not give, or rounds a figure into a new one.
Claude:Met
ChatGPT:Not met
n8n:Met
You decide what to do with each fix
You check this one
Meets:You read each fix, chose to find a source, soften the claim or cut it, and checked the suggested wording against the source yourself.
Does not meet:A fix is used without being read against the source, or a claim marked not in sources is sent as it stood.
Claude:For your read
ChatGPT:For your read
n8n:For your read
Claude: met 8 of 8 (including 1 for your read)
ChatGPT: met 7 of 8 (including 1 for your read)
n8n: met 8 of 8 (including 1 for your read)
Learn from this template
Every claim is compared with words quoted from the source, never with what the AI remembers. A rule checks that each quote really appears in the source, and a claim with no passage is marked as having none instead of being guessed at.
Where to see it: step 2 (Find the passage for each claim); step 3 (Give each claim a verdict).