What problem does it solve? Deciding whether a shared skill actually needs an edit is usually a judgment call based on anecdote. This Skill replaces guesswork with an empirical loop: it mines real session transcripts for recurring friction, proposes a bounded fix, and validates it against held-out cases before anything ships. ## Core Features & Use Cases - Friction Mining: Searches recorded sessions via agentsview MCP tools for skill invocations plus friction phrases, requiring evidence from multiple independent sessions before acting. - Bounded Proposals: Generates narrow diffs touching only the lines implicated by the evidence, still satisfying the skill-authoring-standard rubric. - Validation Gate: Runs held-out eval cases against old and new skill text in parallel, judged by an independent stronger model; accepts only if nothing regresses and at least one case improves. - Use Case: A skill in core/skills/ keeps getting corrected the same way across unrelated sessions. Run this loop to confirm the pattern, produce an evidence-backed diff, and open a reviewable PR — a human always merges. ## Quick Start Run the skill-tuning-loop workflow on the skill that keeps getting corrected to mine session evidence and validate a bounded fix before opening a PR.
