Asking a frontier model “is this any good?” returns its default taste. Hand it a rubric you wrote and the same model returns consistent, comparable, defensible scoring.
Ask Claude Opus 5 to review a draft with no criteria and you get the model's built in disposition: agreeable, fluent, and tilted toward praising whatever it was shown. Ask it to grade the same draft against a rubric you wrote, and the behavior changes completely. Scores become consistent across drafts, across days, and across whoever is running the prompt.
The setup takes two minutes. Before pasting the work in, define three to five criteria that actually matter, and write one line describing what a 1, a 3, and a 5 look like for each. Then ask for a score per criterion, one sentence of evidence quoted directly from the material, and no overall verdict until every criterion has been scored.
This works because it converts a vague aesthetic judgment into a structured extraction task, which is what these models are genuinely reliable at. The model stops guessing at your standards and starts applying stated ones. It also blunts sycophancy: each score has to be defended against your written definition rather than against your feelings.
The payoff compounds on anything you review repeatedly, which is where most professional AI use actually lives: marketing copy, contract triage, code review, vendor comparisons, candidate screens. Write the rubric once, save it, and reuse it. Monday's scores become directly comparable to Friday's, and you can hand the same rubric to a colleague, or to a different model, and line the answers up side by side.
Try it today
- Pick one thing you evaluate over and over.
- Write three to five criteria. For each, describe what a 1, a 3, and a 5 look like.
- Prompt: Score this against the rubric. For each criterion give the score and one sentence of quoted evidence. No overall assessment until every criterion is scored.
- Save the rubric and reuse it every time, so results stay comparable.
