Ask AI to explain exactly what a draft got wrong
"Make it better" gives the assistant room to guess. Show it the source and ask which sentence changed the meaning, and why, and you get a fix you can reuse.
What you'll make
a short list of the mistakes in a draft, each quoting the draft, pointing to the source and naming the kind of error, plus a corrected draft.
What you'll need
the draft and the source it should follow, both of which you’re allowed to share with a work AI chat (ChatGPT, Claude, Copilot or Gemini). Paste them together; an assistant can’t check against a document it can’t see. About ten minutes.
Example result.
| Draft says | Source says | Kind of mistake |
|---|---|---|
| “Q3 travel cost £3,400” | “UK travel: £3,400. Overseas travel is not yet reconciled.” | A subtotal reported as the total |
| “up on Q2 because of the Leeds conference” | Mentions the conference; says nothing about Q2 or cause | A comparison and a cause the source doesn’t give |
1. Ask for mistakes tied to the source
Paste the source and the draft and ask:
Review this draft against the source below. For each mistake that matters, quote the draft phrase, quote the source passage, explain how the meaning changed and suggest a correction.
Name the kind of mistake in a few words, such as "subtotal reported as total" or "cause not in source". Put the ones that would change a reader's decision first.
No overall score and no general praise. Keep the parts of the draft that are right.Why this works: quoting both sides makes every finding checkable, and naming the kind of mistake gives you something to look for next time.
What you should see: a short numbered list, each item with a draft quote, a source quote and a label.
2. Open the source and check each finding
For each item, find the quoted source passage yourself. Does it say what the feedback claims? Assistants can give elegant feedback based on a misreading. Then look at the labels: two findings with the same label are one habit to fix, not two accidents.
Why this works: the source is the authority, not the feedback. And a repeated label tells you what to put in your next prompt.
What you should see: every quote found in the source, and one or two labels that come up more than once.
3. Fix the draft, then hunt for the same mistake
Ask for the correction to the most important finding, and for the same kind of mistake anywhere else in the draft (the repair prompt is below). Then add a line to the instruction you used to write the draft, so it doesn’t happen again: “Report every amount with its qualification, such as ‘UK only’.”
Why this works: correcting the output while keeping the same weak instruction is expensive déjà vu. Fixing the instruction stops the next draft making the same mistake.
What you should see: a corrected draft, any other instances of the same mistake fixed, and one new line in your prompt.
Check the result
- Each finding quotes the draft and the source exactly.
- Each correction is no stronger than the source.
- The same kind of mistake has been looked for across the whole draft.
- Nothing in the correction comes from outside the source.
The usual failure: a qualified figure is reported as the whole. The source says “UK travel: £3,400; overseas not yet reconciled”, the draft says “travel cost £3,400”, and the qualification has vanished because it made the sentence longer. Spot it by checking each figure in the draft for the words around it in the source.
The draft reports "[draft phrase]", but the source says "[source passage]". Correct it so it keeps the qualification. Then check every other figure in the draft for the same mistake, and list each one you change.[draft phrase] and [source passage] are the two quotes from the finding.
Worked example: the travel total that wasn’t
Fictional example. A finance note and the summary drafted from it:
Source: Q3 2026 travel and accommodation
UK travel: £3,400. Overseas travel is not yet reconciled.
Hotels: £1,900, excluding two invoices still to arrive.
The Leeds sales conference took place in August.
Draft summary:
"Q3 travel cost £3,400 and hotels £1,900, up on Q2 because of the Leeds conference."- The first finding was the usual failure. “Q3 travel cost £3,400” dropped “UK” and “overseas not yet reconciled”. The feedback labelled it “subtotal reported as total”.
- The hunt found a second one. The repair prompt, run on that finding, also caught “hotels £1,900”, which leaves out “excluding two invoices still to arrive”. Same label, same fix. The corrected sentence: “UK travel cost £3,400 (overseas not yet reconciled) and hotels £1,900 (two invoices still to arrive).”
- The third error was a different kind. “Up on Q2 because of the Leeds conference” adds a comparison and a cause the source doesn’t contain. It was deleted, not reworded. The label “cause not in source” went into the next prompt alongside the qualification rule.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Ask for the question, not a guess
When the source is genuinely unclear, good feedback names the fact that’s missing and who could supply it, rather than picking an answer. “Ask finance for the overseas travel figure” helps the work continue. A general disclaimer doesn’t.
Turn a repeated label into a checklist
If the same label keeps appearing across drafts, it belongs in a checklist for that kind of work. The task-specific checklist guide shows how, and executive summary source trails covers tracing every claim in a summary back to its source.