JEV GUIDE

Jev prompt examples: Choice, Score and Noul

Define the decision your code needs before writing the question. These editorial examples are synthetic and have not been tested against the live model.

Jev Gallery · Reviewed 2026-09-19

Replace vague goals with testable judgments

“Is this feedback good?” is difficult to evaluate. Separate bug detection, feature area and completeness into distinct judgments.

Question IDs are not passed to the model, according to the documentation. A key named has_steps still needs instructions that define what counts as reproduction steps.

Source: TypeSafe · Questions ↗

Choose the type for the output

Use Choice to select a queue, Score for explicit completeness levels, and Noul for a single condition. The following question fragment needs a state and model to form a full request. Its other option handles inputs outside the defined queues.

JSON

{
  "route": {
    "type": "choice",
    "instructions": "Which queue best matches the main request in `report`?",
    "criteria": {
      "bug": "A feature behaves differently from an explicitly described expectation.",
      "idea": "A request to add or change functionality without a described failure.",
      "other": "Unclear, unrelated, or insufficient information."
    }
  },
  "detail": {
    "type": "score",
    "instructions": "How much reproduction information is explicitly present in `report`?",
    "criteria": [
      "No specific action or outcome.",
      "An action and unexpected outcome are described.",
      "Actions, expected outcome, and environment are described."
    ]
  },
  "has_environment": {
    "type": "noul",
    "instructions": "Does `report` explicitly name a browser, operating system, or application version?"
  }
}

Source: TypeSafe · Questions ↗

Probability is not intensity

A Noul near 0.5 expresses uncertainty about whether a condition holds. It is not a medium degree of that condition. Check explicit evidence for a browser version rather than asking the model to guess; use ordered levels for an intensity judgment.

Source: TypeSafe · Noul ↗

Look for instruction failures

Label expected queues and reasons before running a batch. Keep genuinely ambiguous examples in the evaluation set.

Mixed bug and feature request: define how to choose the main intent.
Negation such as “not a bug”: inspect whether it is misread.
Missing information: retain an other or review path.
Irrelevant instructions inside the input: keep the task bounded and validate outputs in code.
Mixed languages and paraphrases: evaluate each instead of assuming equivalence.

Version the experiment

Store fixtures, questions and expected handling together. Change one question or a related set of criteria at a time, then compare errors. Use suggestion mode before decisions alter data; a model answer is not permission to execute an action.