Turn one AI task into a useful review checklist
“Check it's accurate” is a wish, not a check. A useful checklist says what to compare, against what, what a failure looks like and what to ask the AI to fix.
What you'll make
a checklist of four to six rows for one AI job you repeat, each row naming what to compare, the evidence, a failure example and the fix to ask for.
What you'll need
any work chat assistant (ChatGPT, Claude, Copilot or Gemini), one recurring AI job such as summarising meeting notes, one real or sample output from that job with its source, and about 20 minutes.
Example result.
| Check | Compare against | Fails when |
|---|---|---|
| Each decision keeps its scope and conditions | The decision line in the notes | “Approved” with the “if…” or “for X only” removed |
| Owners appear only where the notes name one | The action lines | An owner is inferred from who spoke most |
| Every number matches the notes | The figure in the notes | A rounded or recalculated figure |
1. Ask for checks that come from this job
Start from the job, not from “accuracy”. A meeting summary goes wrong by dropping conditions and inventing owners. A spreadsheet answer goes wrong on units and blanks. A portrait edit goes wrong by changing the face. Tell the AI the job, what the output is used for and what has gone wrong before.
Create a review checklist of 4–6 rows for this AI job: [job and output].
The output is used for: [who uses it and what they decide].
Mistakes we have seen: [past failures, or "none recorded yet"].
For each row give: what to check, the evidence to compare it against (not the AI's own say-so),
one concrete example of a failure, and the correction to ask for.
Order the rows by how much damage the failure would do. No confidence scores.[job and output] is something like “summarise weekly operations meeting notes into decisions and actions”. The other two brackets give the AI the consequences, which is what decides the order.
Why this works: “independent evidence” is the important column. Asking the same assistant “is your summary right?” is conversation, not checking. Each row has to point at something outside the output.
What you should see: four to six rows you could run in a few minutes, starting with the failure that would cost most.
2. Test the checklist on an output with a known fault
Take one output where you already know something is wrong, or plant one fault yourself: delete a condition, change a number. Run the checklist against it. If the checklist doesn’t catch the fault, the checklist needs work, not the example.
Here are the source notes and an AI summary with one fault I know about.
Run the checklist on the summary row by row. For each row, quote the source line and the summary line you compared and say pass or fail.
If the known fault isn't caught by any row, sharpen or add a row so it would be. Remove any row you couldn't actually check from these two texts.Why this works: a checklist that has never caught anything is untested. One known fault proves at least one row works, and shows which rows are decoration.
What you should see: a pass or fail per row with the two quoted lines, and the known fault caught.
3. Turn a failed row into a precise repair
When a row fails, ask for that one fix, not a fresh rewrite. A full rewrite can quietly break the parts that passed.
Row "[row name]" failed: the output says [output wording], but the source says [source wording].
Correct only that line so it matches the source. Change nothing else, and tell me which line you changed.Paste the two wordings exactly as they appear, so the AI can’t paraphrase its way back to the same mistake.
Why this works: a narrow correction keeps the passing rows passing, so you only re-check what changed.
What you should see: the corrected line, the rest unchanged, and the changed line named.
Check the result
- Every row names evidence outside the AI output: a source line, a file, a known answer.
- Every row has a concrete failure example, not “is inaccurate”.
- The checklist caught the known fault in step 2.
- It fits on one screen. If nobody will run it, it isn’t a checklist.
The usual failure: the AI writes generic rows (“Check for accuracy”, “Ensure clarity”, “Verify completeness”) that sound thorough and can’t fail. Spot it by asking of each row: what exactly would I compare, and what would a fail look like? If you can’t say, the row is generic.
These rows are too generic to fail: [row names].
Rewrite each one for [job and output] so it names the exact evidence to compare against
and one concrete failure example from this kind of work. Delete any row that still can't fail.[row names] are the rows you flagged; [job and output] is the same job description as step 1.
Worked example: a courier decision that lost its condition
Fictional example. Kestrel Freight’s operations lead wants a checklist for the AI summaries of the weekly courier review.
First checklist (generic): “1. Check the summary is accurate. 2. Ensure all key points are covered. 3. Verify the tone is professional.” Run against a summary with a planted fault, it passes everything.
Repair: the repair prompt above, with all three rows named. The rewritten checklist is the table under “Example result”, plus two rows: “open questions stay open” and “dates match the notes”.
Notes: Approved: move north-region parcels to Swiftline, only if the 4-route trial keeps damage under 2%. Trial review 18 Nov (Priya).
Summary: Decision: switch to Swiftline. Priya to review on 18 Nov.Step 2 test with the rewritten checklist: row 1 fails. The summary drops both the scope (“north region”) and the condition (“only if damage stays under 2%”). An unconditional “switch to Swiftline” could get a contract signed for every region. The other rows pass: Priya is named in the notes, and 18 Nov matches.
Repair of the summary: the step 3 prompt, with row 1 named and both wordings pasted. The fixed line reads: “Decision: move north-region parcels to Swiftline if the 4-route trial keeps damage under 2%. Priya reviews the trial on 18 Nov.”
- The generic checklist passed a fault that mattered; the job-specific row caught it in seconds.
- The row compares wording against the notes line, not against what sounds plausible.
- The summary repair named one line and froze the rest, so only row 1 needed re-checking.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Other jobs need different rows
The method is the same; the rows change with the job.
| Job | One row that earns its place |
|---|---|
| Spreadsheet answer | Totals exclude blanks rather than counting them as zero |
| Portrait edit | Face, hairline and neckline match the original photo |
| Customer explanation | Conditions and exclusions match the policy wording exactly |
Keep a known-answer case for analysis checks
For numbers, the best evidence is a small case you worked out yourself before asking the AI. The known-answer guide shows how a tiny, independently calculated case catches errors that a plausible total hides.
Record which checks you actually ran
A checklist is only useful later if someone can see which rows were run on which version. The task run record shows how to log checks, failures and repairs against each output, with unrun checks left visibly pending.