Reusable instruction packages that teach AI coding agents your specific conventions, workflows, and context — without re-explaining them every time.
Skills solve a mechanical repetition problem: every new chat starts from zero. You re-explain that your team uses TanStack Query, not useEffect. That tests go in a specific folder. That the API uses snake_case.
Teams share them through Git — readable, reviewable, diff-able. Source code that affects your software.
Modern implementations support loading only when relevant — saving tokens, preserving the context window.
Skills can bundle supporting scripts, reference documents, and templates the AI can execute or read on demand.
The persistent-memory primitive tools forgot to build.
The reference implementation and its progressive disclosure.
A different philosophy: four activation modes.
Every major coding assistant, side by side.
The predictable shape that makes a skill work.
Ruthless editing, file-size discipline, activation tests.
The brutally honest picture — wins and failures.
Three principles that survive all the hype.