A single score is a conversation-ender in the wrong direction. "This prompt is a 72" invites exactly one reply: says who? The point of a rubric is to make that number impossible to wave away — because it is not one judgment, it is ten small ones you can inspect line by line.
Every Rubrkit audit scores an instruction on the same ten dimensions. They are not arbitrary. Each is a place real instructions fail in production, phrased as a question you can answer with evidence from the text.
The ten
- Objective clarity — does the instruction state the single outcome that defines success, not just a topic or a format?
- Output specification — is the shape of a correct answer described precisely enough that two readers would recognize the same thing as "done"?
- Evaluation criteria — could a second person grade the result the way you would, without asking you what you meant?
- Bounded behavior — does it say what the model must not do, and what to do at the edges?
- Context sufficiency — does the instruction carry everything it needs, or does it assume knowledge that isn't on the page?
- Constraint quality — are the rules specific and checkable, or soft words like "appropriate" and "concise" that mean nothing under pressure?
- Failure handling — is there a defined move for missing input, ambiguity, or the empty case?
- Reusability — could this run on the next case without a rewrite, or is it welded to one example?
- Maintainability — when someone edits it in six weeks, will they understand what each part is load-bearing for?
- Safety — does it guard the things that hurt when they go wrong — PII, destructive actions, irreversible steps?
Why ten and not one
A single number hides where the weakness is. Ten scores turn "make it better" from a mood into a punch list: you see which dimension is soft, you fix that clause, and you watch that specific score move. The overall grade is a rollup — useful for a threshold in CI — but the work happens in the dimensions.
Clear objective. Bounded behavior. Testable result.
Those three refrains are shorthand for the cluster that fails most often. An instruction that nails objective clarity, bounded behavior, and a testable output specification is already past the failure modes that burn the longest runs.
The rubric is the same everywhere
The dimensions you read in a web audit are the dimensions the CLI gates on and the MCP tools return. npx rubrkit audit, rubrkit_start_audit, and the report in the app all score the same specimen the same way. That consistency is the point: a grade only means something if it means the same thing tomorrow, in CI, in your editor, and in the pull request.
An instruction you can only judge by feel is one you can only argue about. Ten dimensions give you something better than an argument — a diff.