Blog / Questions / FIG. 54
Can Jev Write Code? No, and Knowing Why Makes You Better at Using It
Jev cannot write code, essays, or anything else: it's a decision model. What that means, why it's the point, and what developers use it for instead.
No. Jev does not write code, and it never will, because it does not generate text of any kind. Not functions, not essays, not one-line replies. It answers closed questions (a label, a yes/no, a score, a pick-from-list) extremely fast and extremely cheaply, and that's the entire product.
This is the single most common misconception the launch created, so it's worth being precise about the boundary. Generation and decision are different jobs: a chat model like GPT or Claude composes; a decision model rules. Asking Jev to write a function is like asking a referee to play striker; the refusal isn't a bug, it's the role.
What developers actually use it for is the judgment layer around code, and there's more of it than most people notice: PR merge-readiness verdicts (scope, tests present, secrets, description-matches-diff, per the developers guide), flaky-versus-real test triage, log dedup and severity routing, semantically checking CI outputs your assertions can't, and housekeeping inside coding agents, where one cataloged build uses instant keep/drop verdicts for context compaction. The code gets written by you or your frontier model; Jev is the fast second opinion on everything checkable, at reported prices where checking everything is the default.
The pairing, not the replacement, is the pattern: cascade your capable model for generation and reasoning, Jev for the fifty small rulings per task that were never generation problems. Builders who internalize the split report the numbers that fill the directory; builders who ask it for code write the disappointed threads.
Frequently asked questions
Can Jev generate any text at all?
No: no code, prose, summaries, or rewrites. Its outputs are structured answers from closed sets, natively, with probabilities per choice per ecosystem clients.
Can Jev review code?
It can judge checkable properties of a diff (tests touched, scope, secrets, description accuracy) as structured verdicts; it cannot write review prose or reason about architecture. Pre-reviewer, not reviewer: the full pattern.
What should I use for code generation then?
A frontier chat model, with Jev as the verdict layer around it: the cascade that keeps showing up in the directory's strongest builds.
Numbers throughout are as reported by the build authors, not verified by shipwithjev. Code-shaped examples are pseudocode; the official docs live at docs.typesafe.ai.