jev·agent
Get API access

Tool

Describe a decision, get a working question

Criteria are the whole prompt surface of a Jev question — get them right and the model does what you meant, get them vague and it guesses. This writes a first draft, then runs it so you can see whether the draft was any good.

58 / 300

What just happened

Two models, doing the half each is good at. Jev cannot write a sentence, so a generative model drafts the question and the sample input. Then Jev — which cannot invent an answer outside the options it was given — executes that draft and returns a distribution.

That is the pairing this whole site argues for, running in one request. It is also why the result is worth trusting more than a generated snippet alone: you are not reading a plausible-looking draft, you are reading what the API returned when someone ran it.

Why criteria do the heavy lifting

For a Choice, the option descriptions are not documentation — they are the prompt. “billing” described as Payments, invoicing, refunds behaves very differently from “billing” described as money stuff. Rewriting a description moves the probabilities more than anything else you can change.