skill tutorial · agent
Skills for agents that install nothing
Strip the loader away and a skill is two things: a description that matches your brief, and instructions that run. The first works in any chat window; the second becomes a checklist you can execute in ten seconds. No directory required.
Pick the right skill without the whole catalog
Open skills.html or skills.json, filter by category + minutes + group size, read the one-line descriptions. That is the discover step — a browser does it fine.
Paste and prompt
the pattern
1. Open skills/<name>/SKILL.md in the site (raw markdown). 2. Paste the whole file into the chat, prefixed: "Follow this skill. Do not run commands — output the exact commands for me to run, plus the shopping list scaled to 14 people, and the debrief order." 3. The agent answers from the folder's facts, not invention.
What you lose, honestly
- No execution: the curl lines are homework — one paste into a terminal runs the whole prep.
- No standing context: pasting is per-conversation; agents with rules/directories keep the drawer loaded.
- Nothing else: matching, scaling, print lists, agenda math, debrief sequencing all survive the paste — they live in the file.
Make it repeatable
Most chat products have custom instructions or projects. One paragraph there — When I brief a workshop, I mean WorkshopSkill folders: ask which skill matches, then follow its SKILL.md when pasted
— and the paste step becomes routine: paste the file, describe the session, get the prep.
keep going
questions
Does pasting SKILL.md count as using the skill?
Functionally yes — the file is the interface; loaders only automate finding and feeding it.
Will the chat model hallucinate around it?
Less than bare prompting: the folder's metadata and run sheet constrain the answer. Keep asking for the file's facts, not general advice.
Which agents get this natively?
Claude Code, Codex CLI, Qoder and the growing list with skills directories — see the per-agent walkthroughs linked from skills.html.