v0.1.0

First versioned release.

• Put a skill on a diet. The agent measures what the skill loads, moves each fact to the lowest level that still reaches the agent in time, cuts what the model already knows and what only a human needs, and keeps the slim version only if agents do at least as well with it.
• Keep it lean afterward. Rules for every later update stop new lessons and code changes from bringing the bloat back, and a skill that grows more than 10% past its last diet gets another.
• Skills that teach a tool get extra steps: the calls agents make most from your session logs, mismatches between the skill and the tool proved one by one, and a check that warns when the skill drifts from the tool.
• Before/after evals run the same tasks with the old skill, the new one and no skill, on any coding agent. Run a diet right after using the skill and that use becomes one of the tasks. Your agent writes a small adapter for itself, and three short calibration runs check it before every eval.
• Runs unattended. Eval tasks are written before anything is cut, a fresh session settles disputed answers without seeing which version wrote them, and the scorer fails the diet if the slim skill does worse than chance would explain, or if most runs never opened the skill.
• Skills only you start get marked user-only, so their description stops loading in every session.
• Install with pi install, the Claude Code plugin, or npx skills add for Codex, Cursor and other agents.