skill tutorial · agent
Workshop skills in Amp
Amp threads are sharp but stateless between conversations — so the durable pattern is a folder on disk plus one standing instruction. The skill bundle was built for exactly this: plain files, printed commands, no loader to wait for.
Install
a drawer outside any repo
mkdir -p ~/skills && cd ~/skills curl -LO https://workshopskill.com/en/activities/start-stop-continue/files/start-stop-continue.zip unzip start-stop-continue.zip
Teach the thread once
Amp lets you set persistent system/persona instructions — add:
standing instruction
Workshop skills: ~/skills/<name>/SKILL.md (Agent Skills format, from https://workshopskill.com/en/skills.json). Prep briefs: match metadata first, read only the match, run commands as printed, ask before live-form (formlm-cli) steps.
First run
Open a thread in your workshop directory: retro for one team of 6, 45 minutes, tensions under the table
. Amp reads start-stop-continue's SKILL.md — the behavior-not-person rule and the owner-and-date rule are in the facilitation notes it just loaded — fetches the card stock list and debrief, and drafts the agenda. The thread ends with files on disk, not advice in scrollback.
Notes
- New thread, same drawer: the standing instruction carries; nothing to reinstall.
- Long sessions: point at files by path rather than pasting SKILL.md bodies — Amp reads on demand.
- Team usage: keep ~/skills in a shared repo; the bundle directories diff cleanly.
keep going
questions
Does Amp support SKILL.md natively?
Amp grows agent features fast; even with no loader, the folder-plus-instruction pattern gives the same behavior — matching is done by the model, knowledge lives in the file.
Which config file holds the standing instruction?
Amp's persona/system instructions (workspace settings) — one paragraph, done; exact menu names move with versions.
Can it run formlm-cli?
Yes, threads execute shell commands with review; login is interactive, covered by the skill's headless fallback.