What problem does it solve? It automates the end-to-end lifecycle of authoring new skills: detecting reusable patterns in your own work, drafting SKILL.md files, generating test cases, running with-skill vs baseline benchmarks, iterating on failures, optimizing trigger descriptions, and packaging the result. ## Core Features & Use Cases - Pattern Detection and Deduplication: Scans git diffs, attempt history, tool usage, and cross-episode notes to find recurring patterns, then checks existing skills to avoid overlap. - Automated Benchmarking: Generates eval cases with assertions, spawns parallel with-skill and baseline runs, grades outputs via grader/comparator/analyzer agents, and aggregates pass rates, timing, and token deltas. - Description Optimization and Packaging: Runs a trigger-accuracy optimization loop over should-trigger/should-not-trigger queries and packages the validated skill into a distributable .skill file. - Use Case: After noticing you repeatedly write the same data-cleaning helper across tasks, use this Skill to capture it as a tested, benchmarked, packaged skill without manual eval design. ## Quick Start Analyze my recent work for a repeated pattern and turn it into a tested, packaged skill. ## Reference Files The agents/ directory contains grader, comparator, and analyzer subagent instructions; references/schemas.md defines the JSON formats for evals, grading, and benchmarks.
