Writing Great Skills

This is a Chinese translation of Matt Pocock's meta skill for writing better Claude skills. The core idea is that skills exist to squeeze determinism out of randomness by making the process predictable, not the output. It walks you through the model-invoked versus user-invoked tradeoff, which comes down to context load versus cognitive load. The real insight is treating your description as a cost center: every word in a model-invoked description burns tokens on every turn, so you trim ruthlessly or flip it to user-invoked if the agent doesn't need to discover it automatically. If you're building skills or maintaining a collection that's gotten unwieldy, this gives you a clear framework for when to let the model see something and when to hide it behind manual invocation.

Install to Claude Code

npx -y skills add vinvcn/mattpocock-skills-zh-cn --skill writing-great-skills --agent claude-code