Self-Reflection Prompting: Make AI Critique Its Own Outputs

This is Article 3 of 9 in our series: Advanced Prompt Engineering Mastery. Previous: Tree-of-Thoughts and Graph Prompting. Next: Anti-Hallucination Prompting.

The First Draft Is Never the Best Draft

Every skilled writer, engineer, and strategist knows that the first version of anything is a starting point, not a destination. The gap between a first draft and a polished output is not filled by the original act of creation — it is filled by revision: stepping back, seeing what is actually there rather than what you intended, and making deliberate improvements.

Standard AI prompting collapses this process into a single step. You ask, the model answers, you accept or reject. There is no revision loop unless you manually drive one — and most people do not, because they do not know how to structure it systematically.

Self-reflection prompting gives the model an explicit revision loop. Instead of asking it to produce a finished output, you ask it to produce a draft, critique that draft against specific criteria, and then improve it based on its own critique. Recursive self-improvement extends this into multiple iterations, each cycle building on the last. The result is measurably better than a single-pass prompt — and the improvement is not marginal.

In the previous article, we explored how branching reasoning helps AI explore a problem space before committing to an answer. Self-reflection solves a different problem: it improves the quality of an answer after the model has already committed to one. These two techniques are complementary — branch first to find the right direction, then reflect and refine to maximise quality within that direction.

The Research Behind It

The formal study of AI self-reflection emerged from several directions at once. The most influential early framework was Reflexion, introduced by Noah Shinn and colleagues at Northeastern University in a 2023 paper. Reflexion showed that language models could significantly improve their performance on coding, reasoning, and decision-making tasks by storing their own failure signals as verbal feedback and using that feedback in subsequent attempts. On the HumanEval coding benchmark, Reflexion agents reached a 91% pass rate compared to 80% for standard prompting.

A related framework, Self-Refine, introduced by Madaan et al. at Carnegie Mellon University in another 2023 paper, demonstrated that models could generate feedback on their own outputs across diverse tasks — dialogue, code optimisation, math reasoning, and sentiment reversal — and use that feedback to produce consistently better second drafts. Crucially, Self-Refine required no additional training or fine-tuning: the improvement came entirely from prompt structure.

The insight is counterintuitive: you do not need a smarter model to get better outputs. You need a smarter prompting structure that allows the same model to revise its own work.

 

Three Levels of Self-Reflection

Not all self-reflection is equal. There are three distinct levels, each appropriate for different tasks and token budgets.

Level 1 — Single-pass critique. The model produces output, then critiques it within the same response. Fast, low-cost, and appropriate for most everyday tasks. This is the minimum viable version and should be appended to almost every substantive prompt.

Level 2 — Separated critique and revision. Two separate prompts: one to generate, one to critique and revise. The separation matters because the model approaches its own output with fresh distance — it is no longer anchored to the reasoning it used to produce the draft. This yields noticeably sharper critiques than the single-pass version.

Level 3 — Recursive iteration. Multiple cycles of critique and revision, each building on the previous. Two to three cycles is the practical ceiling — beyond that, diminishing returns set in as the model begins defending its prior choices rather than genuinely reconsidering them. Reserve this level for high-stakes outputs: important proposals, final client deliverables, complex technical documents.

Template 1: Single-Pass Critique (Level 1)

Append this block to any prompt where output quality matters. It adds minimal tokens relative to the output you are already generating.

[Your standard prompt here — generate the output as normal]

After producing your response, add a section titled SELF-CRITIQUE.
In this section:
1. List 3 specific weaknesses in what you just wrote. Be precise —
   do not say "could be clearer," name exactly what is unclear and why.
2. List 1 thing that works well and must be preserved in any revision.
3. Write a REVISED VERSION that addresses all 3 weaknesses
   while keeping the identified strength intact.

The instruction to name one strength before revising is deliberate. Without it, models tend to demolish their first output entirely — producing something different rather than something better. Anchoring one strength focuses the revision on targeted improvement.

Template 2: Separated Critique and Revision (Level 2)

Prompt A — Generate:

[Your task description and context]

Produce a complete first draft. Label it DRAFT 1.
Do not self-edit while writing — produce your best immediate response.

Prompt B — Critique and revise (paste Draft 1 in full):

Below is a draft that needs critical review. Approach it as a
demanding editor who has not seen it before.

DRAFT 1:
[paste draft here]

ORIGINAL GOAL: [restate what this output was meant to accomplish]
TARGET AUDIENCE: [who will read or use this]

CRITIQUE PROTOCOL:
Score the draft on each dimension from 1 to 5:
- Accuracy: Does it get facts and logic right?
- Clarity: Can the target audience follow it without effort?
- Completeness: Does it cover everything the goal requires?
- Tone: Does the register match the audience and purpose?
- Structure: Does the order of information serve the reader?

For any dimension scoring below 4, identify the specific passage
responsible and explain precisely what is wrong.

REVISION:
Produce DRAFT 2 — a complete rewrite addressing every identified
weakness. Write the full output, do not summarise or truncate.

Template 3: Recursive Self-Improvement (Level 3)

Two full critique-and-revise cycles in one prompt. Use this for deliverables where quality has a real impact: a client proposal, a key article, a complex technical brief.

You will produce a high-quality output through structured iteration.

TASK: [full task description]
GOAL: [what success looks like]
AUDIENCE: [who this is for]
QUALITY STANDARD: [specific criteria — e.g. "a non-specialist must
understand every sentence without asking a follow-up question"]

CYCLE 1:
Write DRAFT 1. Then critique it using these questions:
- What is the weakest section, and why?
- Where does the logic or structure break down?
- What would a critical reader object to first?
- What is missing that the goal requires?
Write DRAFT 2 that fixes every identified problem.

CYCLE 2:
Read DRAFT 2 as if you are the target audience encountering it
for the first time. Ask:
- Is the opening compelling enough to continue reading?
- Is every technical term explained or avoidable?
- Does the ending deliver on the promise of the opening?
- What single change would most improve this draft?
Write DRAFT 3 incorporating your answers.
Label it FINAL OUTPUT.

Scoring Criteria: The Variable That Changes Everything

The quality of a self-reflection prompt is determined almost entirely by the quality of the criteria. Generic criteria produce generic critiques. Specific criteria produce actionable ones.

Use Case	Weak Criteria	Strong Criteria
Article or blog post	“Is it clear and engaging?”	“Would a reader who skims only the opening paragraph and the headings understand the core argument?”
Translation	“Is it accurate and natural?”	“Does any passage read like a translation rather than original writing in the target language? List each one.”
Code review	“Are there any bugs?”	“What are the three most likely failure points under unexpected input? What breaks first under load?”
Client proposal	“Is it persuasive?”	“What objection would a skeptical client raise after reading paragraph 2? Does the proposal preemptively answer it?”
Research summary	“Is it comprehensive?”	“What is the strongest counterargument to the main claim? Is it acknowledged?”

The pattern is consistent: move from an adjective to a specific test a real reader would apply. Adjectives give the model permission to declare success. Tests force it to verify.

What Self-Reflection Cannot Fix

Self-reflection improves structure, clarity, tone, and internal logic. It cannot fix factual errors the model does not know are errors — the model cannot critique what it does not know is wrong. For factual reliability, pair self-reflection with the verification techniques in the next article. (See our article: Anti-Hallucination Prompting.)

It also cannot fix a fundamentally wrong approach. If the first draft solves the wrong problem, recursive refinement produces a polished wrong answer. This is why Tree-of-Thoughts and self-reflection work best in sequence: branch first to confirm the right direction, then reflect and refine to maximise quality within it.

Model Notes for 2026

• Claude 3.7 Sonnet: Produces the most useful self-critiques — structural problems, not just surface ones, with revisions that address root causes. Enable extended thinking for Level 3 tasks.
• GPT-4o: Strong at Levels 1 and 2. Tends toward diplomatic feedback — sharpen it with “identify the single weakest sentence and explain specifically why it fails.”
• o3 / o4-mini: Internal deliberation makes self-reflection partly redundant on logical tasks. Use it for outputs with a subjective quality dimension — writing, strategy, communication.
• Open-source models (Llama 3.3, Mistral Large): Use the separated Level 2 approach. Single-pass critique in one prompt tends toward vague feedback. Two separate calls consistently produce better results.

Common Mistakes

Asking for critique without criteria. “Critique your response” produces observations like “could be more concise.” That is useless. Criteria turn it into “the third paragraph contains two claims that contradict each other — here is what needs to change.”

Accepting the first critique as complete. Models front-load obvious problems and bury more significant ones. Follow up: “Beyond what you identified, what is the next most important problem?”

Running too many cycles. After three iterations, models begin preserving previous choices rather than reconsidering them. Stop at three and accept that diminishing returns have arrived.

Not specifying what to preserve. Unconstrained revision often produces something different rather than better. Name explicitly what should not change — a tone, a specific example, the structure of the conclusion.

Exercises

1. Baseline comparison: Take any AI output you accepted in the past week. Paste it into a new conversation with the Level 1 critique block appended. Compare the revision to the original and note whether the critique surfaced problems you had consciously noticed.
2. Criteria calibration: Write three critique criteria for a type of output you produce regularly — one adjective-based (weak), one general test (medium), one specific reader scenario (strong). Run all three on the same draft and compare the feedback quality.
3. Recursive drill: Take a short piece of writing and run it through two full critique-and-revision cycles using Template 3. Read the final version alongside the first draft and note the distance travelled.

Next in the series: Article 4 — Anti-Hallucination Prompting: Self-Consistency, Chain-of-Verification and Reliable AI Outputs.

---

References

1. Shinn, N. et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. Northeastern University. arxiv.org/abs/2303.11366
2. Madaan, A. et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. Carnegie Mellon University. arxiv.org/abs/2303.17651
3. Zy Yazan — Tree-of-Thoughts and Graph Prompting. zyyazan.sy
4. Zy Yazan — Anti-Hallucination Prompting. zyyazan.sy