The Rubric Trick: Make It Grade Its Own Draft Before Handing It Over

The gap between a model's first draft and its third is large, and you normally pay for it in round trips. This closes most of it in one shot.

At the end of the prompt, after the actual task, add:

`
Before you answer, write a 4-item rubric for what a great answer to this
looks like. Then draft the answer. Then score your own draft against each
rubric item out of 5 and note what is missing. Then give me the revised
answer only - not the rubric, not the scores, not the first draft.
`

Why it beats "think step by step": step-by-step reasoning improves the path to an answer. A rubric improves the target. The model writes down what "good" means before it has an answer to defend, so the standard is not retrofitted to justify whatever it happened to produce.

Why it has to write the rubric itself. I tried supplying my own. It performs worse. A supplied rubric gets treated as a checklist to satisfy minimally; a self-authored one gets treated as intent.

Why the last line matters so much. Without "revised answer only" you get four screens of scaffolding and have to scroll to find the deliverable. The reasoning still happens; you just do not have to read it.

Where it does not help: anything with one correct answer. Grading your own arithmetic is not a rubric, it is a second guess. Save it for work where quality is a judgement call - writing, design critique, architecture options, naming.

Costs maybe 40% more tokens. Saves two round trips. I run it on anything I would otherwise have edited by hand.