jev-pref

# jev-pref

**Turn your preferences from AGENTS.md into a fast, Jev-powered AI linter.**

`jev-pref` lets you define project-specific semantic rules, run them against
code changes with [Jev](https://docs.typesafe.ai), and feed the results back to
your coding agent.

## Quick start

Tell your coding agent:

```text
Run `npx jev-pref setup` and follow the instructions it prints.
```

That's it.

`setup` inspects the repository and teaches the agent how to configure
`jev-pref`. The agent explains the system, asks you a few questions, helps
translate your preferences into useful semantic checks, and adds persistent
instructions to `AGENTS.md`, `CLAUDE.md`, or wherever you choose.

No special agent integration or global installation is required. Node.js 20+
and `npx` are enough.

## What jev-pref is for

Think:

```text
TypeScript  → type invariants
ESLint      → syntax and static rules
tests       → behavioral invariants
jev-pref    → semantic project rules
```

Instead of asking an AI "is this code good?", you define what matters, Jev
classifies the evidence in the change, `jev-pref` maps the result to an
outcome, and your coding agent acts on it. The full boundary — what Jev may
and may not judge — lives in [docs/principles.md](./docs/principles.md).

Good questions name externally defined, evidence-grounded checks:

```text
Does this diff introduce new mutable module-level state?
```

```text
Does this change remove or rename an existing exported symbol?
```

```text
Classify the API impact:

- none
- additive
- behavioral
- breaking
```

Poor questions ask Jev to invent a standard of quality:

```text
Is this good architecture? Is this code clean? Are these tests sufficient?
```

If you cannot explain what visible evidence would make an answer true, the
rule needs more