Stop Asking AI to Critically Self-Check

The short version

“Critically self-check this” sounds responsible.

In practice, it often creates a tiny reward hack: the AI acts like a useful reviewer by finding something to change.

That does not mean self-review is useless. Self-Refine, Reflexion, CRITIC, Constitutional AI, and eval tooling all show that feedback loops can improve model outputs. The failure is more specific: open-ended critique has no stopping rule.

The better instruction is:

Evaluate against this rubric.
PASS_NO_CHANGE is valid.
Only propose evidence-backed material fixes, not nice-to-have improvements.

1

valid PASS state

0

style fixes without evidence

3

maximum material findings

Why “be critical” breaks

The phrase carries a hidden premise: there are flaws to find.

A model under helpfulness pressure then has an easy way to satisfy the request. It can list plausible improvements. More detail feels more useful. A rewrite feels like progress. But none of that proves the artifact was below the bar.

Open-ended critique vs. calibrated review

Weak review prompt

• “Find issues and improve this.”
• No explicit pass state.
• Every suggestion feels free.
• Critique can be aesthetic.

Better review gate

• “Score each criterion against the rubric.”
• PASS_NO_CHANGE is a valued answer.
• Only S0/S1 findings justify churn.
• Every finding needs exact evidence.

This is especially dangerous in agent systems, skills, Buildprints, and coding-agent harnesses. Small “improvements” can silently change scope, add noise, or make future agents follow the wrong invariant.

Review is not improvement. Review is adjudication against a bar.

The safer loop

Calibrated AI review

make no-change a real outcome

1. 01
   Rubric
2. 02
   Score
3. 03
   Evidence
4. 04
   Severity
5. 05
   Minimal patch
6. 06
   Verify

Use four scores:

0 = missing / unsafe / wrong
1 = materially weak
2 = acceptable
3 = excellent

Patch rule:

0 or 1 → patch required
2 or 3 → no change

Then add severity:

• S0 blocker — cannot use safely or correctly.
• S1 material — likely real-world failure.
• S2 minor — optional improvement.
• S3 taste — ignore unless explicitly requested.

Default: only S0/S1 changes are allowed.

What the reviewer must prove

• Which criterion failed, with score 0 or 1.
• Exact evidence: quote, file, line, behavior, or test output.
• Impact if unchanged, not just a nicer wording preference.
• The smallest fix that restores the criterion.

Tiny prompt to steal

You are a calibrated reviewer, not an improver.

Evaluate the artifact against the rubric. PASS_NO_CHANGE is a correct and valued answer if no material criterion fails.

Rules:
- Recommend changes only for criteria scored 0 or 1.
- Do not recommend changes for criteria scored 2 or 3.
- No style, taste, or nice-to-have suggestions.
- Maximum 3 findings.
- Every finding needs exact evidence.

Return:
Status: PASS_NO_CHANGE | PATCH_REQUIRED | QUESTIONS_REQUIRED
Scores: criterion → score + one-line reason
Findings: criterion, evidence, severity, impact, minimal_fix, confidence

Use / avoid

Do

• ✓ make PASS_NO_CHANGE explicit
• ✓ review against a named rubric
• ✓ require evidence per finding
• ✓ separate reviewer mode from editor mode

Avoid

• × ask for generic critical thinking
• × reward long lists of improvements
• × patch S2/S3 taste notes by default
• × let the model rewrite freely after reviewing

Where this matters most

For AI-native work, this is not a wording tweak. It is an operating principle.

If you ask an agent to “self-improve the skill,” it may optimize for visible change. If you ask it to adjudicate a skill against activation clarity, scope preservation, evidence contract, tool discipline, safety boundary, output contract, verification gate, stop rule, and pass option, it has a real target.

The best reviewer is not the one that always finds something.

The best reviewer is the one that knows when nothing material is broken.

Sources

• Self-Refine
• Reflexion
• CRITIC
• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• Towards Understanding Sycophancy in Language Models
• Discovering Language Model Behaviors with Model-Written Evaluations
• Constitutional AI
• Anthropic: Define success criteria
• Anthropic: Evaluation Tool
• LangSmith Evaluation