Asking an AI to “make this better” usually produces a smoother but blander version of your text, often with changes you didn’t want. A more reliable approach is to separate two steps: first critique against explicit criteria, then improve based on that critique. You stay in control of what changes, and the results are more targeted.

The prompt

I’m going to share a draft. First, act as a demanding editor and critique it against these criteria:

  1. [criterion 1 — e.g., clarity of the main point]
  2. [criterion 2 — e.g., evidence for each claim]
  3. [criterion 3 — e.g., appropriate tone for the audience]
  4. [criterion 4 — e.g., length and concision]

For each criterion, give a score from 1 to 5, quote the specific passages that need work, and explain why.

Do not rewrite anything yet. After the critique, ask me which issues I want you to fix.

Audience: [who will read this] Goal: [what the text should achieve]

Draft: [paste text]

Then reply with the issues you want addressed, for example: “Fix issues 1 and 3, keep my examples and voice, don’t add new claims.”

Why it works

  • Explicit criteria focus the feedback on what matters to you rather than generic polishing.
  • Quoting passages makes the critique concrete and checkable.
  • Separating critique from rewriting lets you reject bad suggestions before they’re applied.
  • Iterative refinement with feedback reflects techniques studied in research, such as Self-Refine, where models improve outputs by generating feedback on their own drafts, and Constitutional AI, where models critique and revise responses against written principles.

Example criteria by use case

Work email: main request stated in the first two sentences; polite but direct tone; clear deadline; under 150 words.

Essay or article: clear thesis; logical structure; evidence for each claim; strong opening; no repetition.

Cover letter: specific to the company; concrete achievements; matches the job requirements; authentic voice.

Code: correctness; edge cases; readability; security issues; test coverage. (Ask for a critique before refactoring.)

Presentation slides: one idea per slide; readable text; logical flow; strong conclusion.

Variations

Role variation: “Critique this as [a skeptical investor / a first-time user / a hiring manager].”

Rubric variation: paste an official rubric or style guide and ask for a critique against it.

Comparison variation: share two versions and ask which better meets the criteria, and why.

Self-review variation: after the AI produces something, ask it to critique its own output with the same criteria before you use it.

Tips

  • Keep criteria to between three and six; more becomes unfocused.
  • Ask for specific quotes to avoid vague feedback.
  • Tell the AI what to preserve: your voice, examples, structure or facts.
  • Don’t accept every suggestion — the critique is input to your judgment.

Common mistakes

  • Asking for critique and rewrite in one step, which hides what changed.
  • Using vague criteria like “make it better”.
  • Letting the AI add facts or claims you can’t verify.

A worked example

Suppose you’ve written a project update for your manager. Using criteria of clarity, evidence, tone and length, the AI might score clarity 3/5, quoting a paragraph that buries the main risk at the end, and evidence 2/5, noting that “we’re making good progress” has no figures. You then reply: “Fix clarity and evidence: move the risk to the top, and use the numbers from my notes below. Keep my tone.” The result is a targeted revision you control — not a generic rewrite.

What to watch next

  • Writing tools with built-in rubric-based feedback.
  • Research on self-critique and model reliability.
  • Education policies on AI feedback for students.
  • AI review features in code and document editors.

Key takeaways

  • Separate critique from rewriting to keep control over changes.
  • Use three to six explicit criteria tailored to your audience and goal.
  • Ask for scores and quoted passages to make feedback concrete.
  • Specify what to preserve when requesting improvements.
  • The approach reflects research on self-feedback and critique-and-revise methods.

Frequently asked questions

Does this work with any AI assistant?

Yes. The structure works with any capable chat model.

Can I use it for code?

Yes. Use criteria such as correctness, edge cases, readability and security, and ask for a critique before any refactoring.

Is AI feedback reliable?

It is useful but not infallible. Treat it as a second opinion and apply your own judgment.