Jev Skill — Development skill for Claude Code

Jev skill for Claude Code, Codex and other coding agents

**We ran Jev more than 19,000 times. This is the guide we wish we had before call one.** The [research tally](reference/research-scope.md) records 19,367 successful Jev API responses and 90.4M metered input tokens in 4 days of experiments. The useful part is what survived those tests: 5 copyable recipes, a CLI, failures, and the conditions under which each result held.

Start with [**The Ultimate Guide to Jev (After 19,367 Calls and 90M Tokens)**](https://marcodicesare-dev.github.io/jev-skill/) if you want the full story, an honest decision map and a first workflow you can run. The [Markdown version](GUIDE.md) lives here in the repo.

[Jev](https://docs.typesafe.ai/) takes a `state` and focused questions and returns typed answers: **Choice** (one option), **Score** (a described scale), or **Noul** (yes/no). It does not write prose. Your code decides what to do with its answers. This repository adds a standard-library Python CLI, 5 tested question sets, and measurements from real work. It is an independent community project, [not TypeSafe's official skill](https://github.com/typesafe-ai/skills).

Install the Jev agent skill

git clone https://github.com/marcodicesare-dev/jev-skill.git ~/.claude/skills/jev  # Claude Code
git clone https://github.com/marcodicesare-dev/jev-skill.git ~/.agents/skills/jev  # Codex and others
export TYPESAFE_API_KEY=...  # set locally; never paste a real key into a prompt or commit

Then ask your agent: **“Use the Jev skill to find one repeated, bounded decision in this workflow. Show me the question, the state, and a small labelled test before changing the workflow.”** Python 3 and an API key are enough to use the CLI; there are no package dependencies. You can also install [TypeSafe's official skill](https://github.com/typesafe-ai/skills) with `npx skills add typesafe-ai/skills --skill typesafe-ai`; use its live documentation for current provider behavior.

Fir