Six ideas, one per screen. Each is a thing to point at while you explain. The drama is in the watching, so take the silences.
Four moves: grasp the idea, build one, place it among the other tools and share it, then fix it when it misbehaves.
A skill is a folder of instructions. Claude holds only its one-line description until your request matches, then it loads the rest.
Personal skills (~/.claude/skills) follow you across every project; project skills (.claude/skills) ship in the repo for the whole team. The rule of thumb: if you find yourself repeating an instruction, that's a skill waiting to be written.
Create a directory, write SKILL.md, restart. Claude loads every skill's name and description at startup, then matches your request to them.
Claude loads only names and descriptions at startup, then asks you to confirm before pulling the full skill into context. If two skills share a name, priority decides, so keep names distinct.
Two fields are required, two optional. Past about 500 lines, move detail into supporting files Claude opens only when it needs them.
allowed-tools fences a skill (say Read, Grep, Glob, Bash only) for read-only or sensitive work. And tell Claude to run a script, not read it: the source never enters context, only the result does.
Skills load on demand when your request matches. The other four solve different problems. Pick by when each one fires.
Reach for a skill when knowledge should apply automatically the moment the topic comes up. CLAUDE.md is always-on standards; subagents isolate delegated work; hooks fire on events; MCP brings in outside tools. Most setups use several at once.
Commit to the repo for your team, publish a plugin for the community, or push enterprise-managed settings for the whole organisation.
One gotcha: subagents start with a clean context, so they don't inherit your skills. List them in a custom agent's skills: field. Built-in agents (Explore, Plan, Verify) can't use skills at all.
Run the validator first. Then match the symptom to its fix. Most "won't trigger" problems are just the description.
Test the exact phrases you'd type ("why is this slow", "make this faster") and add any that fail to the description. That single fix solves most triggering problems. Run claude --debug to see loading errors by name.
A folder and a SKILL.md turn a repeated instruction into automatic, shareable expertise: built lean, placed among the other tools, fixed by its description.
tutto.one/praxis · daniel@tutto.one