AI Self-Rating Improvement Loop
Make AI critique and upgrade its own work before you accept it
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 98%
Manage AI work as a strong manager would manage a skilled employee whose craft may exceed the manager's own. After the AI produces a deliverable, ask it to score the result from zero to one hundred. Then ask what is wrong, incomplete, or preventing a perfect score. This turns latent quality criteria into an explicit improvement list. Finally, instruct the AI to implement the improvements it just identified and review the revision against that list. The loop separates production, diagnosis, and correction, which reduces the chance that the first plausible answer is accepted as finished. Unlike human feedback, the loop can be repeated without concern for fatigue, competing workload, or the emotional effect of immediately asking for another pass.
Origin
Joshua Wöhle adapted a management technique—asking skilled employees how they would improve their own work—for repeated use with AI assistants.
Core principles
- 01The producer often knows its own domain gaps better than the manager
- 02Explicit scoring forces quality criteria into view
- 03Criticism is useful only when followed by a revision
- 04AI can iterate without the emotional and workload costs of human feedback
How to run it
- 1
Produce the first version
Give the AI the task and enough context to create a complete initial deliverable.
Watch out Do not mistake completion for quality.
- 2
Force a score
Ask the AI to rate its work on a scale from zero to one hundred.
Pro tip Ask it to judge the output against the actual objective, not presentation alone.
- 3
Expose the gap
Ask what is wrong with the work and what changes would bring it to one hundred.
Pro tip Request concrete defects and corrections rather than general reservations.
Watch out The model's critique is not a substitute for factual verification.
- 4
Apply the critique
Tell the AI to make all valid improvements it identified, then return the revised work.
- 5
Check the revision
Compare the new version with the defect list and confirm that each material gap was addressed.
Pro tip Escalate unresolved factual or expert questions to an independent check.
In the wild
An AI drafts a client proposal and rates it 78 out of 100. It identifies weak differentiation, missing evidence, and an unclear next step. The user tells it to implement those corrections and then checks the revised proposal against the three gaps.
→ A plausible first draft becomes a more persuasive and complete proposal.
Common mistakes
Stopping at the score
A numerical rating has no value unless the AI explains the gap and revises the work.
Treating self-review as proof
AI can miss the same flaw twice, so factual claims and high-stakes outputs still need independent verification.
Is it for you?
Best for
It is best for substantial drafts, analyses, proposals, and other outputs that benefit from explicit quality iteration.
Not ideal for
It is not ideal when correctness requires external facts, testing, or expert review that the model cannot supply itself.
From the transcript
“On a scale from 0 to 100, how would you rate that you just did that job?”
“What is what is wrong with it? How would you get it to 100 out of 100?”
“Okay, now go and do it.”
From the episode
Joshua Wöhle: Turn AI Into Your Competitive Advantage and 50X Your Productivity
Joshua Wöhle