The critique loop: make it improve its own work

The one idea

The first answer is a draft. The pros never stop there — they run a critique loop:

1. Get the draft
2. Ask the model to critique it against a standard
3. Ask it to revise using its own critique

"Now review your draft as a skeptical executive with thirty seconds. What would make them stop reading? List the three biggest problems, then rewrite fixing them."

The revision after a critique is consistently, noticeably better than the first draft.

Why this works

Generating and evaluating are different tasks. When the model generates, it commits to early choices and follows them through. When you ask it to evaluate, it reads the text fresh — the same pattern-matching now hunts for problems instead of next words. You are getting two different capabilities out of one tool by asking twice.

Variations that earn their keep

• Named standard: "Critique this against: clear ask in the first line, under 150 words, no jargon."
• Borrowed judge: combine with lesson 4 — critique in the role of your actual audience
• Pre-mortem: "This proposal failed and we are doing the post-mortem. What went wrong?" — surfaces risks a normal review misses
• Stop condition: "Keep revising until you cannot find a meaningful improvement, then show the final version." One prompt, several internal passes

Try it now

Take any AI draft from this week. Run one critique-and-revise pass with a named standard. Compare the two versions side by side.

Your win today

You never have to settle for the first answer again — and the upgrade costs one extra prompt. This loop is the foundation of working with AI rather than just querying it.