lean-jev-usage

# lean-jev One skill, one job: when a task turns on a judgment rather than on writing something, hand that judgment to a model built for judgments — and frame it the way that model needs it framed. ## What it gives you `/lean-jev:ask-jev`, or simply asking to have Jev score something, routes a decision to Jev, TypeSafe's System One model, through `scripts/jev.py`. What comes back is typed: a probability for a yes/no, a position on a rubric you wrote, a pick from a fixed set with the distribution behind it. Numbers only, 70–500 ms, a fraction of a cent. The framing is the work. Jev reads a question literally, answers each one in isolation, and explains nothing, so a question that bundles two factors returns a confident number about neither. The skill holds the rules that keep a question answerable — one factor per question, the right type for the job, state cut down to what the question needs, every number reported next to the question that produced it. ## Ad-hoc questions and recipes **Ad-hoc** covers the one-off: score these candidates, rate each of these tests, does this clear the bar. The general guidelines are enough — the skill picks the type, splits what needs splitting, and batches the lot into a single call. **Recipes** are for questions that keep coming back. A recipe is a written decomposition of one high-level question into Jev-shaped sub-questions, with the criteria spelled out and an honest note on what it leaves unanswered. The plugin ships one, `references/evaluating-a-test.md`, for judging whether a test is worth its lines. `references/writing-a-recipe.md` covers adding your own, including the honesty rules that keep a recipe from promising more than it measures. Recipes are where our own vocabulary accumulates: what "severe" means here, what counts as worth raising, which house pattern a file is being held to. ## Why hand the judgment over at all Three properties the agent's own read lacks: - **No drift down a list.** Every question in a call is evaluated independently and in parallel, so item 40 gets the same attention as item 1. - **No stake in the work.** An agent judging code it wrote an hour ago is a compromised reviewer. - **Consistency.** The same rubric means the same thing next week, which is what makes a threshold worth tuning. ## When to keep it Choosing between architectures or tools, anything needing several hops of reasoning or simulated consequences, arithmetic, counting, dates. That is System Two work — keep it in the agent, or in code. A single judgment with nothing at stake is also cheaper to just answer. ## Requirements The `evaluate` binary installed and a `TYPESAFE_API_KEY` exported; the plugin ships the launcher only. Whatever goes in `state` leaves the machine. Both are covered in the plugin `README.md`, and in `skills/ask-jev/references/setup.md` for when the tool turns out to be missing mid-session.