How to Fix an AI First Draft Without Starting Over
Keep what works, describe the expected result, actual result, and gap, then run one focused revision you can review against the original task.
When an AI first draft misses the mark, do not immediately restart the conversation or discard the entire result. Protect what is already correct, then state the target, the visible gap, the evidence, the requested change, and the completion check. Revision is not another prompt lottery. It is a controlled change.
Suppose AI organizes an event notice. The date and location are correct, and the tone is acceptable, but the reply deadline is missing. You answer, “This is wrong. Rewrite it.” The next version adds the deadline but changes the correct time.
The problem is not that AI cannot revise. “This is wrong” does not say what to preserve or where the real gap is. Treat revision as a small acceptance test, not as another draw from the model.
Turn a vague reaction into an observable difference
“Too ordinary,” “not professional enough,” and “it feels wrong” are real reactions, but they are not yet executable feedback. Ask what you would observe if the next version were better.
| Vague reaction | Observable gap |
|---|---|
| Too long | The draft is 420 words; it needs to be under 200 |
| It does not sound like an internal reminder | It includes company background and formal announcement language, but the recipients know one another |
| It misses the key point | The source contains a reply deadline, but the result does not |
| The classification is confusing | Action items and background are mixed; they need separate sections |
Once you can state what exists now and what should be different, the AI has a chance to make a local correction.
A controlled revision needs five parts
OpenAI recommends defining “Done when” and reviewing and validating the result before accepting it. Anthropic likewise puts explicit success criteria and empirical tests before prompt improvement. OpenAI: Codex best practices Anthropic: Prompt engineering overview
The official guidance supports defining success and checking results. The five fields below are my teaching template for turning that principle into a revision exercise. They are not an official provider template.
- Target: What was the result supposed to accomplish?
- Gap: Which specific part of the first draft fails?
- Evidence: Which source sentence, rule, or example supports your judgment?
- Change: What should be added, removed, rewritten, or reordered this time?
- Completion check: What will you see when this revision is complete?
For a missing reply deadline, write:
The date, time, location, and preparation fields are correct. Keep them unchanged.
Revision needed: the summary is missing the reply deadline.
Source evidence: “Please reply to Aaron by noon on September 9.”
Add a “Reply deadline” field with “Noon on September 9.”
Do not change the other fields.
After the revision, list the fields you actually changed so I can review them.This feedback protects the correct material and limits the revision to one field.
Some failures require missing material, not different wording
If the first draft failed because the model lacked audience context, classification examples, or an output format, add the missing material. “Try to sound more professional” cannot recreate a company style that you never supplied.
If the problem is a nonexistent citation, a bad calculation, or a statement outside the source, do not ask the model to confirm itself. Ask it to show its basis, then verify against the original source yourself. Better prompting can improve task alignment, but it cannot turn uncertain evidence into reliable evidence.
Check both the fix and the damage around it
Many people inspect only whether the original error disappeared. You also need to confirm that correct content was not broken.
| Check | Question | Result |
|---|---|---|
| Requested change | Was the missing reply deadline added? | |
| Protected content | Are the date, time, and location still correct? | |
| New side effects | Did the revision introduce information not found in the source? |
If one error is fixed but two correct fields are damaged, the revision is not complete. Treat the damage as a new visible gap and make another small correction rather than adding ten vague requests at once.
Know when starting over is the better choice
Local revision is not always superior. Rebuild the task when:
- the goal has changed rather than merely narrowed;
- the first draft used the wrong source material;
- the structure is fundamentally unsuitable and little is worth preserving;
- several rounds of correction have produced contradictory requirements.
Even then, do not restart blindly. Save the confirmed requirements, useful source material, good examples, and the reason the first attempt failed. A clean restart should rebuild a reliable foundation, not repeat the lottery.
Practice: complete one traceable revision
Choose one low-risk AI draft and one clearly observable gap.
One-round revision record
Original target:
Correct content that must be preserved:
Specific gap to fix:
Evidence:
Requested change:
How I will know it is fixed:
Did the second version fix the requested issue?
Did it damage correct content?
What still needs review?The goal is not a perfect final document. It is one revision with evidence that you can inspect later.
The next question is different: even when a result follows your format and instructions, are its facts, citations, inferences, and recommendations reliable enough to use? That requires a separate verification method.