Claude Auto Effort — Development skill for Claude Code

claude-auto-effort

A [Claude Code mod](https://code.claude.com/docs/en/plugins/mods/overview) that sets the effort for each prompt you send. [Jev](https://docs.typesafe.ai), TypeSafe's System One model, rates how much careful reasoning the request needs, in about 200 ms before the turn starts. This is the Claude Code port of [pi-auto-effort](https://github.com/justmytwospence/pi-auto-effort), with the same questions and policy.

Score	Request	Level
0	a lookup, trivial answer, or mechanical change	low
1	a routine single-file change or clear question	medium
2	a multi-step change, debugging with clear symptoms, design in a known pattern	high
3	subtle design, cross-cutting refactor, hard debugging, correctness-critical work	xhigh

Jev sees the request, the two before it, the last answer, and what the last run did (tool errors, files edited, tool calls).

**Smoothing.** A running average `e = 0.5·score + 0.5·e_prev` moves the level up when it is at least 0.6 above it, and down when at least 0.6 below and the level has held for 2 prompts. A confident (>= 0.6) score 1.5 or more above the level jumps straight to it. A go-ahead ("yes", "continue", "do it") keeps the level.

**Bounds.** The session's own effort (`/effort`, `effortLevel` in settings) is the ceiling; `low` is the floor. A numeric budget, or a model without effort, is left alone.

**Cache-safe.** The level is decided at `turn.start` and applied to every model request of that turn (`turn.step`), so tool follow-ups never change it. Subagent loops are left alone. Without `TYPESAFE_API_KEY`, or when Jev fails or takes over 1.5 s, the level holds.

Commands

• /auto-effort or /auto-effort status: current level, ceiling, average, last decision.
• /auto-effort on, /auto-effort off: for this session; off sends the session's own effort.

The status line under the prompt shows `effort: high (auto)`. Each decision is written to the debug log (`claude --deb