# Stop Asking AI to Critically Self-Check

Canonical URL: https://huecki.com/en/blog/ai-self-check-always-finds-something/
Markdown URL: https://huecki.com/en/blog/ai-self-check-always-finds-something.md
Language: English
Published: 2026-05-21
Updated: 2026-05-21
Author: Dominic Hückmann
Topic: AI Agent Workflows
- Agent topics: Context Engineering, Agent Evals
- Tags: AI Agents, Prompt Engineering, Evals, AI Safety, Developer Workflow
Content status: field-note

## Summary

Open-ended instructions like “critically self-check this” accidentally reward the model for producing criticism. The fix is not less review. It is calibrated review: explicit criteria, PASS_NO_CHANGE, evidence per finding, severity thresholds, and a tiny change budget.

## Description

Why AI reviewers always find something, and the safer replacement: rubric-based review where PASS_NO_CHANGE is a valid outcome.

## Body

## 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:

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

## 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.

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

Use four scores:

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

Patch rule:

```txt
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.

## Tiny prompt to steal

```txt
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
```

## 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](https://arxiv.org/abs/2303.17651)
- [Reflexion](https://arxiv.org/abs/2303.11366)
- [CRITIC](https://arxiv.org/abs/2305.11738)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://arxiv.org/abs/2306.05685)
- [Towards Understanding Sycophancy in Language Models](https://arxiv.org/abs/2310.13548)
- [Discovering Language Model Behaviors with Model-Written Evaluations](https://arxiv.org/abs/2212.09251)
- [Constitutional AI](https://arxiv.org/abs/2212.08073)
- [Anthropic: Define success criteria](https://docs.anthropic.com/en/docs/test-and-evaluate/define-success)
- [Anthropic: Evaluation Tool](https://docs.anthropic.com/en/docs/test-and-evaluate/eval-tool)
- [LangSmith Evaluation](https://docs.langchain.com/langsmith/evaluation)

## FAQ

### Why does AI self-review often over-edit?

Because instructions like critically self-check imply that defects exist and make a list of improvements look more helpful than a calibrated PASS.

### Should teams stop using AI review?

No. They should replace open-ended critique with rubric-based adjudication, evidence requirements, severity thresholds, and a valid PASS_NO_CHANGE state.

### What is the simplest replacement prompt?

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