Can a model really judge its own work?

Against explicit criteria, usefully often — checking a draft against a stated rule is an easier task than producing the draft. Against no criteria, no: it will simply praise itself.

Why exactly three weak points?

A cap forces prioritisation. Ask for “all the problems” and you get fifteen, mostly trivial, with no signal about which one to fix first.

Prompt foundations

Critique then revise

A two-pass prompt: get a draft, then hand over your real criteria and make the model find its three weakest points before it rewrites anything.

Use this when a draft is 70% right and you can feel what is wrong but cannot articulate the fix.

The pattern

Copy this and replace everything in [SQUARE BRACKETS] with your own specifics.

Prompt template

Here is the draft: [PASTE THE DRAFT, or say "your previous answer"].

Judge it against these criteria, which matter in this order:
1. [MOST IMPORTANT — usually about the audience or the decision being made]
2. [SECOND]
3. [THIRD]

First, list the three weakest points in the draft, quoting the exact line and naming which criterion it fails. Be specific and do not be polite about it.

Then rewrite, fixing only those three things. Do not change anything that already worked.

Why each part is there

Criteria in priority order: Without ranked criteria the model optimises for generic “quality”, which in practice means more words and more adjectives. Ranking tells it what to sacrifice.
Quoting the exact line: Forces the critique to be grounded in your text instead of abstract advice like “could be more engaging”. A quote is checkable; a vibe is not.
Only those three things: Prevents the total rewrite where fixing the intro silently destroys the one paragraph you liked. Bounded edits are reviewable edits.

Before and after

The lazy version

Make this better.

What goes wrong: “Better” against no criteria means longer and more florid. You get a draft that is different rather than improved, and the part you were happy with is gone.

The pattern, filled in

Worked example

Here is the draft: [the project README you just wrote].

Judge it against these criteria, which matter in this order:
1. A recruiter with 40 seconds and no ML background can tell what this project does and why it was hard.
2. Every claim about results is backed by a number that appears in the repo.
3. Someone can run it from a clean machine using only these instructions.

First, list the three weakest points in the draft, quoting the exact line and naming which criterion it fails. Be specific and do not be polite about it.

Then rewrite, fixing only those three things. Do not change anything that already worked.

What that buys you: Instead of a vaguely nicer README you get a named defect: “Achieves strong performance fails criterion 2 — no number.” Three precise faults, three precise fixes, and the sections that worked survive untouched.

Now do it on your own work

Take something you have written and are slightly unhappy with. Write your three criteria FIRST, before you show the model anything. Then run the pattern and compare its three weakest points against your own gut feeling.

Check yourself — every box should be true

You wrote the criteria before seeing any critique.
Each of the model's three points quotes a real line from your draft.
You disagreed with at least one point and can explain why your judgement beats its criterion.
The rewrite changed those three things and demonstrably left the rest alone.

Then explain it back

Writing the criteria was the hard part. What did ranking them force you to decide about who this piece is actually for?

Write two or three sentences in your own words. If you cannot, you copied a prompt rather than learning a pattern — which is the whole difference this library is trying to make.

Does this depend on which AI tool I use?

Works in one long chat or as two separate prompts. Two separate prompts are often stronger: a fresh context does not see the draft as “mine” and tends to critique harder. Where a tool exposes a temperature setting, a lower value gives a more literal, less inventive rewrite — usually what you want here.

Why it actually works

Ranked criteria plus a scored draft is what a loss function is: one explicit statement of what counts as better, which is the only thing that makes improvement measurable.

See what “better” looks like as a number →

Prompting well is a skill you can pick up in a week. Understanding what is happening underneath is the 18-month curriculum — and it is what stays useful when the tools change.

FAQ

Can a model really judge its own work?: Against explicit criteria, usefully often — checking a draft against a stated rule is an easier task than producing the draft. Against no criteria, no: it will simply praise itself.
Why exactly three weak points?: A cap forces prioritisation. Ask for “all the problems” and you get fifteen, mostly trivial, with no signal about which one to fix first.

Next patterns

The Prompt Patterns Library is a separate, optional track — it does not change the 78-week curriculum or your roadmap. See what the curriculum actually covers →