How Prompt Scoring Works

Most prompt tools tell you a prompt is bad. Very few tell you why.

Promptimize grades every prompt against five criteria and returns a score out of 100, a breakdown of which criteria fell short, and a rewritten version. This page explains what each criterion measures, why it changes the answer you get back, and what the difference looks like in practice.

The five criteria aren't arbitrary. They map to the five things a language model has to infer when your prompt doesn't supply them — and inference is where quality goes to die. Every point a prompt loses is a decision you left to the model.

The five criteria

1. Role & Identity

What it measures: whether the prompt tells the model who it should be while answering.

A model with no assigned role answers as a generalist. That's the blandest version of what it can do. Assigning a role narrows the distribution of possible answers toward the expertise you actually want, and it does so before the model writes a single word.

The second version gets you specifics — notice periods, manager relationships, compensation bands — because you told it which shelf of knowledge to pull from.

Common failure: naming a topic instead of a role. "Write about marketing" is a subject. "You are a demand-gen marketer at a Series B startup" is an identity.

2. Context Sufficiency

What it measures: whether the prompt supplies the background the model would otherwise have to guess at.

This is the single most common reason a prompt underperforms. You know your situation; the model knows only what you typed. Every unstated fact becomes an assumption, and the model will pick the most generic one available.

The first gets you a template. The second gets you something you can send.

Common failure: assuming a follow-up conversation will fix it. It usually can — but the score measures whether the first response will be usable, and a well-contexted prompt skips two rounds of clarifying questions.

3. Output Specification

What it measures: whether the prompt defines the shape of the answer — format, length, and structure.

Unspecified output means the model chooses, and its default choice is prose of medium length with headings you didn't ask for. If you need a table, a numbered list, JSON, or exactly 200 words, saying so costs one clause and saves an entire revision cycle.

Common failure: specifying content but not container. "Cover pricing and support" tells the model what to include. It doesn't say whether you want three paragraphs or three rows.

4. Constraints

What it measures: whether the prompt bounds tone, scope, audience, and what to leave out.

Constraints are the criterion people skip most and regret most. They're what stops a model from hedging, padding, adding disclaimers, or drifting into adjacent topics you didn't ask about. Negative constraints — what not to do — are especially underused and especially effective.

Common failure: treating constraints as optional politeness. They're the difference between an answer you edit and an answer you send.

5. Reasoning Guidance

What it measures: whether the prompt tells the model how to approach the problem, not just what the problem is.

For anything involving analysis, comparison, or judgment, telling the model to work through it in a particular order measurably changes the output. Asking it to consider alternatives before recommending one, or to state its assumptions first, produces reasoning you can actually check.

Common failure: using it everywhere. Reasoning guidance matters enormously for analysis and barely at all for "rewrite this sentence." The grader scores this criterion the same way regardless, so a genuinely simple prompt still loses points here — chasing 100 on a one-line request is not worth it.

How the criteria combine

Each criterion is worth up to 20 points, so the five add up to 100 — there is no hidden weighting, and the breakdown you see is the arithmetic. Two ceilings shape the result: a prompt missing role, context, or output specification is capped at 70 no matter how strong the rest is, and missing any one of the five caps the total at 85. Scoring above 85 requires all five present and working.

A high score doesn't mean a prompt is clever. It means the model has to guess about very little. That is usually the same thing.

Why prompts score low

In rough order of frequency:

1. No role assigned. The most common single miss, and the cheapest to fix.
2. Context the writer has and the model doesn't. You know the audience, the history, the constraints. None of it is in the prompt.
3. No output format. You get prose when you wanted a table.
4. No length bound. You get 900 words when you wanted 200.
5. Stacked questions. Four unrelated asks in one prompt; the model answers the first well and the rest badly.
6. Politeness treated as instruction. "Please do a good job" is not a constraint.

Frequently asked questions