skill-tuning-looplisted

# Skill Tuning Loop A skill earning its place in the lacquer (per `skill-authoring-standard`) is a judgment call today. This loop adds a second, empirical bar: does a proposed edit measurably reduce the friction real sessions hit, without regressing what already works? Run it via the `skill-tuning-loop` Workflow — never edit a skill's behavior-affecting content based on mined evidence without going through the validation gate below. ## When to run this - A skill keeps needing the same correction across unrelated sessions — that's a signal worth mining, not a one-off to shrug off. - Before accepting an automation-proposed edit (see the continuous-tuning design this skill's workflow implements) — the gate is what makes an unattended proposal safe to merge. - Never on every session, and never on a single session's evidence — one grumpy correction is noise, not a pattern. Require independent evidence from multiple sessions before proposing a change. ## The loop (implemented in the `skill-tuning-loop` Workflow) 1. **Mine.** Search recorded sessions (via the agentsview MCP tools) for invocations of the target skill plus friction phrases ("no that's wrong", "actually", "don't do that"). Pull the surrounding messages for anything that looks skill-relevant. 2. **Reflect.** Have an agent read the current `SKILL.md` against that evidence and name concrete, recurring patterns — what the skill says (or omits) and what went wrong because of it. Require ≥2-3 in