What problem does it solve? Writing effective skills for Claude is hard to get right on the first try — descriptions may undertrigger, instructions may be ambiguous, and there is no built-in way to measure whether a skill actually improves outputs. This Skill provides a structured draft-test-review-iterate workflow with quantitative benchmarking so skill quality is measured rather than guessed. ## Core Features & Use Cases - Guided Skill Authoring: Captures intent through interview questions, then drafts SKILL.md with proper frontmatter, progressive disclosure structure, and writing patterns. - Parallel Evaluation Runs: Spawns with-skill and baseline subagent runs for each test prompt, captures timing and token data, and grades outputs against assertions. - Benchmark Aggregation & Review Viewer: Aggregates grading results into benchmark.json/benchmark.md with pass rates and deltas, and generates an HTML review interface for qualitative human feedback. - Description Optimization: Runs an automated loop that tests trigger accuracy against should-trigger/should-not-trigger queries and iteratively improves the skill description. - Use Case: You want to build a skill that generates weekly status reports. This Skill helps you draft it, create realistic test prompts, run side-by-side comparisons against a no-skill baseline, review outputs in a browser viewer, and refine until the benchmark shows consistent improvement. ## Quick Start Ask the AI to create a new skill for your workflow, for example: help me create a skill that converts meeting notes into structured action items. ## Environment Notes The full workflow (subagents, browser viewer, description optimization) is designed for Claude Code. On Claude.ai or Cowork, adaptations apply — no subagents on Claude.ai, and static HTML output instead of a server in Cowork.
