Your AI agent hands you broken work with full confidence and never tells you.
Even at 95% accuracy per step, a 10-step agent is wrong ~40% of the time.
5 self-review prompts that make it catch its own mistakes before they cascade 👇
Even at 95% accuracy per step, a 10-step agent is wrong ~40% of the time.
5 self-review prompts that make it catch its own mistakes before they cascade 👇
First, the mistake to avoid.
Asking an agent "did you make any mistakes?" barely works. It re-reads its own logic, agrees with itself, and hands you the same answer with more confidence.
What works is forcing a different angle. That means a fresh critic role, an explicit checklist, and a check against the source of truth (the spec, the data, the tests), not just vibes.
Every prompt below is built that way.
Asking an agent "did you make any mistakes?" barely works. It re-reads its own logic, agrees with itself, and hands you the same answer with more confidence.
What works is forcing a different angle. That means a fresh critic role, an explicit checklist, and a check against the source of truth (the spec, the data, the tests), not just vibes.
Every prompt below is built that way.
Prompt 1: Code review (before you merge)
You just wrote this code. Now switch roles. You're a senior engineer seeing it for the first time, and you're skeptical it works.
Go through it against this checklist.
- What breaks on empty, null, or unexpected input?
- Where could this fail silently?
- Any security or data-loss risk?
- Does it do what the task actually asked, or just what was easy?
List every issue, rank by severity, fix the ones that matter, and tell me what you changed.
You just wrote this code. Now switch roles. You're a senior engineer seeing it for the first time, and you're skeptical it works.
Go through it against this checklist.
- What breaks on empty, null, or unexpected input?
- Where could this fail silently?
- Any security or data-loss risk?
- Does it do what the task actually asked, or just what was easy?
List every issue, rank by severity, fix the ones that matter, and tell me what you changed.
Prompt 2: Article review (before you publish)
Re-read the draft as a hostile editor who's looking for a reason to reject it.
Hold it to these standards.
- Is every factual claim traceable to a source in your context? Flag any that aren't.
- Does the opening line earn the next one, or can it be cut?
- Where does it pad, repeat, or hedge?
- Where would a skeptical reader stop reading?
Mark the single weakest paragraph, then rewrite the draft to fix what you found.
Re-read the draft as a hostile editor who's looking for a reason to reject it.
Hold it to these standards.
- Is every factual claim traceable to a source in your context? Flag any that aren't.
- Does the opening line earn the next one, or can it be cut?
- Where does it pad, repeat, or hedge?
- Where would a skeptical reader stop reading?
Mark the single weakest paragraph, then rewrite the draft to fix what you found.
Prompt 3: Analysis review (before you conclude)
Before you commit to this conclusion, argue against it.
Run these checks first.
- What assumption, if it's wrong, breaks the whole thing?
- What's the strongest alternative explanation you dismissed?
- Which claims are proven by the data vs. ones you inferred?
- Where are you most likely overconfident?
Then give me the conclusion again, downgraded to only what you can actually defend.
Before you commit to this conclusion, argue against it.
Run these checks first.
- What assumption, if it's wrong, breaks the whole thing?
- What's the strongest alternative explanation you dismissed?
- Which claims are proven by the data vs. ones you inferred?
- Where are you most likely overconfident?
Then give me the conclusion again, downgraded to only what you can actually defend.
Prompt 4: Plan review (before you execute)
Don't run this plan yet. Pressure-test it first.
Walk it step by step.
- Which step is most likely to fail, and what happens to every step after it?
- What are you assuming is true that you haven't verified?
- Which actions are hard or impossible to undo?
- What's the cheapest check you can run right now to de-risk the riskiest step?
Revise the plan so a failure in any single step can't quietly corrupt the rest.
Don't run this plan yet. Pressure-test it first.
Walk it step by step.
- Which step is most likely to fail, and what happens to every step after it?
- What are you assuming is true that you haven't verified?
- Which actions are hard or impossible to undo?
- What's the cheapest check you can run right now to de-risk the riskiest step?
Revise the plan so a failure in any single step can't quietly corrupt the rest.
Prompt 5: Output review (before you hand it back)
Compare your output to the ORIGINAL request, line by line.
Run through the following.
- Did you answer every part of what was actually asked, or drift to a nearby question?
- Recompute any numbers or counts independently. Do they still match?
- What did you assume the user wanted that they never actually said?
- If you had to bet money this is correct, which part would you worry about?
Flag every gap, fix it, then tell me exactly what you verified.
Compare your output to the ORIGINAL request, line by line.
Run through the following.
- Did you answer every part of what was actually asked, or drift to a nearby question?
- Recompute any numbers or counts independently. Do they still match?
- What did you assume the user wanted that they never actually said?
- If you had to bet money this is correct, which part would you worry about?
Flag every gap, fix it, then tell me exactly what you verified.
The pattern under all 5 is the same.
Models are better at spotting errors than at avoiding them. So you generate first, then force a separate, skeptical pass against a real standard.
Bolt one of these onto your agent as a verification step before it writes a file, sends a message, or closes a task.
That single step is the gap between a demo that works and an agent you can trust in production.
Models are better at spotting errors than at avoiding them. So you generate first, then force a separate, skeptical pass against a real standard.
Bolt one of these onto your agent as a verification step before it writes a file, sends a message, or closes a task.
That single step is the gap between a demo that works and an agent you can trust in production.
If you found these useful, also check this out.
Turn Claude into 20+ different specialists for marketing & business.
Install real expertise, not just prompts.
Get my Claude skills bundle 👇
linktr.ee/alex_prompter
Turn Claude into 20+ different specialists for marketing & business.
Install real expertise, not just prompts.
Get my Claude skills bundle 👇
linktr.ee/alex_prompter
• • •
Missing some Tweet in this thread? You can try to
force a refresh
