If you're writing skills for Claude, this is the reference manual. It's all about making skills predictable, meaning the agent takes the same process each run, not necessarily producing identical output. The core tradeoffs are clear: model-invoked skills cost context load but fire autonomously, user-invoked skills cost cognitive load but stay out of the window. The vocabulary matters: leading words, completion criteria, context pointers, progressive disclosure. It teaches you to collapse duplication into single sources of truth and hunt no-ops sentence by sentence. Written as flat reference rather than steps, which makes sense since you consult it on demand while editing, not follow it sequentially.
npx -y skills add mattpocock/skills --skill writing-great-skills --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
supercent-io/skills-template
supercent-io/skills-template
huangjia2019/claude-code-engineering
reactjs/react.dev
reactjs/react.dev