Chapter 3 · Lesson 1

Critique Before You Rewrite

Diagnose why an answer is weak before asking AI to generate another version.

7 min · Reviewed July 2026

Scenario

You keep asking AI to “make it better”, but each new version changes the wording without fixing the real problem.

Name the failure precisely

Separate factual, structural, tonal, and strategic problems. An answer can be accurate but poorly prioritised, clear but unsupported, or polished but aimed at the wrong audience. Different failures require different corrections.

Evaluate against explicit criteria

Give the model a short review rubric such as accuracy, relevance, evidence, clarity, and actionability. Ask it to identify the weakest area and cite the part of the draft that caused the judgment. You still decide whether the critique is right.

Revise only after diagnosis

Once the problem is clear, request a focused revision and preserve what already works. This prevents a full rewrite from introducing new errors or discarding useful details merely because one paragraph was weak.

Build the mental model

Revision improves when critique identifies the failure, evidence, and smallest useful change.

Use a fixed review sequence

Check factual support first, then task relevance, structure, clarity, and tone. Fixing style before evidence can make a weak answer more persuasive without making it more correct.

Require evidence for the critique

Ask the reviewer to quote or identify the passage behind each finding. This makes vague criticism challengeable and prevents a model from inventing problems.

Worked example

A proposal sounds polished but does not make a decision easy.

Weak approach

Make this proposal stronger and more convincing.

Improved approach

Do not rewrite yet. Review the proposal in this order: factual support, relevance to the decision, prioritisation, and clarity. For each issue, quote the passage, explain the consequence, and propose the smallest correction. Then revise only the two highest-impact issues while preserving supported content.

Why it works

The workflow separates diagnosis from generation, orders review by consequence, and limits revision so good material is not casually replaced.

A practical five-line critique

For each issue record: criterion, evidence from the draft, consequence for the reader, proposed correction, and whether a person must verify it.

If reviewers disagree, inspect the criterion before rewriting. The disagreement may reveal that “good” was never defined for the task.

Apply it to your work

Diagnose a draft you already have

Choose three review criteria, identify evidence for the weakest one, and request one bounded revision.

Starter templateReview for [criteria]. Quote the evidence. Name the consequence. Recommend the smallest correction. Do not rewrite until I confirm.

prompt choice

Which follow-up is most likely to improve a weak proposal?

Compare the prompts closely. Look for useful context, constraints, and a clear output.

Takeaway

Diagnose the failure before generating the next version.