AI Prompt Evaluator

Why evaluate a prompt instead of rewriting it

Most "prompt improver" tools hand you back a different prompt and hope it is better. An AI prompt evaluator takes the opposite side of the deal: it tells you what is wrong with the one you wrote — the subject is vague, there is no style direction, the lighting is unspecified — and how confident it is in each judgment. You keep authorship of the fix. That matters most where every generation costs money: on an image platform with 29 models, a prompt that fails on clarity fails 29 times in a row.

This is exactly the job the Jev model was built for. Jev is a structured evaluation model: it does not generate text, it answers typed questions about a state — with probabilities, not prose. Underneath this page, your prompt is the state and the questions are fixed (clear? missing what? how specific?), so every evaluation returns the same shape you can act on immediately.

Three answer types, one Jev evaluation

The same three types scale far beyond prompts — the demo above also includes Jev's classic presets for support triage, refund review, account risk and lead qualification, so you can see how one structured evaluation model covers very different workflows.

The Jev model at a glance

UnitPic calls Jev through the same Cloudflare AI Gateway REST API it uses for third-party image models — one integration pattern for every model.

Calling Jev yourself

Jev is available to any Cloudflare Workers AI account. You define the questions; the model returns calibrated answers:

const response = await env.AI.run('typesafe/jev', {
  state: 'a nice picture of a city',
  questions: {
    is_clear: {
      type: 'noul',
      instructions: 'Is this image prompt clear enough?',
      criteria: { true: 'Unambiguous', false: 'Vague wording' },
    },
    specificity: {
      type: 'score',
      instructions: 'How specific is the prompt?',
      criteria: ['Vague', 'Reasonable', 'Highly specific'],
    },
  },
})
// → { answers: { is_clear: { noul: 0.08 },
//               specificity: { score: 0.3, ... } } }

AI prompt evaluator FAQ