Turn a classifier's structured-output schema into decision questions
Paste the structured-output schema, tool definition or bare JSON Schema your classifier uses. Each field becomes a decision question or is marked keep in LLM, and you get the request in the format you pick.
Runs in your browser. Nothing you paste is uploaded.
curl
How the answers come back
| Question | OpenAI | Gemini | Jev / Clef |
|---|---|---|---|
| Yes/no | {type: "predicate", name, probability} | {type: "noul", noul (0.01–0.99), confidence} | {type: "noul", noul} |
| Pick one | {type: "choice", name, choice, probabilities: [{value, probability}], confidence} | {type: "choice", choice, probabilities: {option: p}, confidence} | {type: "choice", choice, confidence, probabilities: {…}} |
| Ordered rubric | {type: "score", name, score, probabilities: [{value, label, probability}], confidence} | {type: "score", score, legend: {"0": …}, probabilities: {"0": p}, confidence} | {type: "score", score, confidence, legend, probabilities} |
| Refusal | {type: "refusal", name} (name is null for an unnamed question) | not documented | not documented |
| Container | array, in question order | map keyed by question id | map keyed by question name (Clef wraps the response in result: unconfirmed) |
17 warning types. Most link to the provider's docs; format facts come from the sources listed on each format pair page (last checked 2026-11-19).
How to use it
- Paste your schema. An OpenAI
response_formatortext.format, an Anthropic toolinput_schema, a function tool’sparametersand a bare JSON Schema all work. - Check the mapping table. Tick Ordered? for an enum whose values have an order (low, medium, high) to ask a score question instead of a choice.
- Rewrite any placeholder question, then copy the request for the API you picked.
How the mapping works
Each leaf of the schema is mapped the same way every time:
| Schema field | Becomes |
|---|---|
string with an enum of 2 or more values | Pick one (or score, when you tick Ordered?) |
boolean | Yes/no |
integer with minimum/maximum within 10 levels, or an integer enum | Score, one level per value |
array of enum values | One yes/no per value |
number named like confidence, probability or certainty | Dropped: the API returns it |
| Anything else | Keep in an LLM step |
Question keys are the field path joined with _, then sanitized for the target API. Questions without a description get a placeholder you should rewrite. The input is set to <paste the text you classify>.
FAQ
Which classifier fields can a decision API answer?
Enums (pick one), booleans (yes/no), small integer ranges or integer enums (score), and arrays of enum values (one yes/no per value). Objects with properties are walked, so their enum, boolean and integer fields become questions too. Free text, other numbers and objects without properties need a generative step.
What happens to my `confidence` field?
A number field named like confidence, probability or certainty is dropped (an integer or enum one still becomes a question). Decision APIs return probabilities and confidence for every answer, so you don’t ask for them.
What about nullable enums?
For a pick-one question, the converter adds an other option (unless the enum already has one). A field you mark Ordered? becomes a score with no other level. OpenAI recommends a fallback option such as "other" when your categories don’t cover every possible input.