Ask an AI to answer a question, then immediately ask it to score that answer out of 10 — and explain where it lost points. Do this a few times and something shifts: you stop reading AI outputs passively and start treating them like drafts that need a reviewer. Which is what they are.
This isn’t a trick. It’s a prompting pattern that surfaces the model’s own uncertainty in a form you can actually act on.
Why It Works
When you ask a model to generate something and score it in the same pass, it defaults to confident output. Generation doesn’t pause to flag weak spots — it just moves forward. Adding a second step specifically about quality evaluation gives the model permission to switch modes. Critique and generation are genuinely different tasks, and models handle them differently when you separate them explicitly.
The score itself matters less than the explanation. A model that says “7/10 — I’m less confident about the third point because I don’t have reliable data past 2023” is giving you something actionable. A model that just produces a polished answer with no self-assessment is giving you the same uncertainty, wrapped in prose that hides it.
Claude tends to be more forthcoming here than ChatGPT. Ask Claude to score a nuanced answer it just gave, and it’ll often genuinely identify where it was speculating versus where it was on solid ground. Ask ChatGPT the same thing and you sometimes get a 9/10 with flattering commentary. That’s not universal — it varies by task type and how you phrase the request — but it’s a real difference worth knowing if you’re choosing which model to use for high-stakes drafts.
How to Actually Do It
There are two ways to set this up. The first is sequential: get your answer, then follow up with a separate message asking the model to critique it. Simple, works fine, but the model is now in a slightly different context and may be more generous toward its own previous output.
The second is to build the scoring into the original prompt. Something like:
Answer the following question as well as you can. Then, on a separate line, rate your answer from 1-10 and list the two or three specific things you were least confident about or that a knowledgeable reader might push back on.
This works better because the model handles both tasks in one pass — the critique is generated while the generation is still fresh. You also avoid the politeness drift that can happen when a model is reviewing something it already committed to in an earlier message.
The one thing to avoid: asking for a score without asking for the reasoning. A bare number is useless. “6/10” with no explanation tells you nothing you can do anything with.
What the Scores Actually Tell You
Low scores on factual tasks — research questions, technical explanations, anything involving specific numbers or dates — are usually honest signals. The model knows when it’s in uncertain territory on facts, and the self-assessment tends to map onto the places where you should go verify things independently.
High scores on creative or writing tasks are less reliable. A model can confidently rate its own email draft as excellent when the real problem is that it doesn’t match your voice at all. That’s a gap between “technically correct” and “actually useful for you,” and the model has no access to what you sound like. For writing tasks, the scoring is more useful for catching structural problems (“I could have addressed the main objection more directly”) than for judging overall quality.
One thing I’ve noticed running this pattern repeatedly: the critique often contains the rewrite. If a model says “my explanation of X in paragraph two was too vague and didn’t give a concrete example,” you can just tell it to fix exactly that. The gap between a mediocre output and a good one is frequently just that nobody asked the model to name the specific weak spots before moving on.
A Real Limitation Worth Naming
Self-assessment has a ceiling. A model can’t reliably catch errors it doesn’t know it made. If it’s confidently wrong about a fact — not uncertain, but actually mistaken and unaware of it — the self-scoring step won’t surface that. The model scores its confident wrongness the same way it scores confident correctness, because from the inside, they feel identical.
This is where the pattern is most dangerous to misuse. If you treat a high self-score as a substitute for verification on anything checkable — a specific claim, a statistic, a protocol name, a release date — you’ll eventually get burned. The scoring step is a filter for uncertainty, not a fact-checker. Those are different things.
Use it to triage: low-scoring answers need work before you trust them, high-scoring factual answers still need spot-checking on anything specific and important, high-scoring creative answers need your own judgment about fit. The score tells you where to spend your review time, not whether to bother reviewing.
When This Is Worth Adding to Your Workflow
Not every prompt needs a self-assessment step. For quick, low-stakes tasks — brainstorming a list, rephrasing a sentence, generating options — the overhead isn’t worth it. You’ll read it and know fast enough whether it’s usable.
The pattern earns its place when the output matters enough that a bad one has real consequences. Research you’ll act on. A draft you’ll send to someone. An explanation you’ll use to make a decision. Anything where you’d be genuinely annoyed if a flaw slipped through.
It’s also useful early in a new workflow, when you don’t yet know what the failure modes look like. Asking for self-critique a few times teaches you where the model tends to be weakest on your specific task — information you’d otherwise discover by making actual mistakes.
Try it once on something you’d normally just read and use: get your answer, then ask the model what a smart, skeptical expert would find wrong with it. That second answer is often more useful than the first one was.
