skill-tunerlisted

A reference for writing any document an agent consumes — a skill, an `AGENTS.md`, a `CLAUDE.md`, or a doc reached by a pointer from one of those. It descends from the `writing-for-agents` doctrine and keeps its shape. **If you landed here wanting a document improved rather than a reference to read: the user-invoked `tune` skill is the loop — `/skill-tuner:tune <path>` in Claude Code; clients without slash commands surface it as the `tune` skill directly.** It finds this doctrine's defects in a real document, adversarially verifies each one, and reports only what survived — which is a different job from the rules below, and the one most people arrive wanting. The rules below are what that loop fixes *toward*, and what to apply when writing something new. The bundled runner behind it — probe, routing-parity, paired compare — spends tokens, so it starts only on the human's explicit instruction, and its reports land under `reports/` as inputs to a human decision. Every rule carries exactly one evidence tag: `[research]` (grounded in published findings about model behaviour), `[measured]` (grounded in an eval or red-team run against this harness), or `[craft]` (compositional advice with no direct experiment behind it yet). Read the tag as a confidence label, not decoration — a `[craft]` rule is worth following, not worth defending past the point a `[measured]` rule contradicts it. The falsifier for each rule, the citations behind the `[research]` tags, and the runs behind the `[