This is meta tooling for building other skills on the platform. It walks you through the full cycle: capturing intent, writing the SKILL.md with proper frontmatter, creating test cases, running evals, and iterating based on results. The workflow is solid: draft, test with actual prompts, review quantitative and qualitative feedback, rewrite, repeat. It even includes a separate optimizer for skill descriptions to improve triggering accuracy, which matters since Claude apparently undertriggers skills by default. Most useful when you're building something with objectively verifiable outputs where you actually want benchmarks. The progressive disclosure system it enforces (metadata, body, bundled resources) keeps context manageable while letting you attach unlimited reference material.
npx skills add https://github.com/bytedance/deer-flow --skill skill-creator