Self-Criticism Loop
Get the model to critique its own answer, then implement its own critique
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 3
- Confidence
- 93%
After an LLM produces an answer, ask it to check or criticize its own response, then tell it to implement that criticism and rewrite. Schulhoff frames this as a near-free performance boost: generate, self-criticize, self-improve. It is bounded — one to three iterations, not infinite.
Origin
One of Sander Schulhoff's five favorite techniques from The Prompt Report; belongs to a family of self-criticism prompting methods.
Core principles
- 01A model can surface flaws in its own output when explicitly asked to
- 02Separating criticism from revision produces better rewrites than asking for a better answer directly
- 03Diminishing (and eventually degrading) returns — cap the loop
How to run it
- 1
Get the naive answer
Ask your question normally and let the model produce its first response.
- 2
Ask it to check and criticize itself
Prompt: 'Can you go and check your response — confirm it's correct or offer yourself some criticism.' It returns a list of critiques.
- 3
Tell it to implement the critique
Respond 'great criticism, now go ahead and implement that' and it rewrites the solution incorporating its own feedback.
Pro tip Run the loop one to three times at most.
Watch out Beyond a few passes the model 'goes crazy' — it starts changing things when nothing is left to improve.
In the wild
Schulhoff routinely solves a problem, asks the model to criticize its own output, praises the criticism, then asks it to apply the fixes — repeating one to three times.
→ A free quality boost on answers in situations where the task admits self-checking.
Common mistakes
Looping infinitely
Running the self-criticism cycle too many times, at which point the model degrades rather than improves.
Is it for you?
Best for
Any user wanting a cheap quality lift on reasoning or drafting tasks that the model can plausibly self-evaluate
Not ideal for
Tasks with no objective self-check, or when latency/cost of extra passes outweighs the gain
From the transcript
“can you go and check your response, you know, like confirm that's correct or offer yourself some criticism”
“So, it outputs something, you get it to criticize itself, and then to improve itself”
“I'll do it like one to three times sometimes, but not beyond that”
From the episode
AI prompt engineering in 2025: What works and what doesn’t
Sander Schulhoff (Learn Prompting, HackAPrompt)