skill tutorial · scenario
Half an hour, one session, fully prepped
The claim is testable: you have 30 minutes between now and a real session, an agent with a shell, and a browser. This is the whole loop on a stopwatch — brief to bundle to print order — with the minutes written down so you can check the promise.
Minutes 0-5 — discover, don't browse
filter instead of search
curl -s https://workshopskill.com/en/skills.json | jq -r ' .skills[] | select(.category=="communication" and (.group_size|test("^[0-9]+$")) and (.group_size|tonumber) <= 2 and (.duration_minutes|tonumber) <= 20) | .name' # or in the chat, to the agent: "12 people, 90 minutes, no projector, opener + core + close, one measurable core. Propose three candidates with agenda math."
Minutes 5-15 — install and read the winner
one bundle per chosen activity
cd ~/workshop-2026-10-07 for s in blind-portrait exit-ticket; do curl -LO https://workshopskill.com/en/activities/$s/files/$s.zip unzip -o $s.zip done # SKILL.md's metadata answers the real questions: # props -> shopping list, duration -> agenda math, setting -> no-screen OK
Minutes 15-25 — run the prep commands
The agent executes the printed curl steps (props, paper form, debrief, run sheet into one folder). Then the two human gates: print order — one prop set per pair, one answer sheet per person, run sheet for you — and the live-measure decision: assessment.json on paper (zero accounts) or the formlm-cli lane (needs the login moment, produces the delta report).
Minutes 25-30 — rehearse the failure modes
the questions worth asking out loud
> what goes wrong in round A of blind portrait, and what does > the folder tell me to do about it? the debrief file names three failure modes with rescues; the facilitation notes say the silence is the design, not you failing to facilitate — the prep is done when you can say that in the room without checking the sheet
keep going
questions
Does this really fit 30 minutes?
The timer is the point of the guide — with a skills loader installed, discover+install is under five minutes of curl and reading; printing overlaps with the rest. First run, give yourself 45.
What if I want more than the drawer offers?
Name the constraint in the brief: category, headcount, minutes, setting — the catalog filters are the same fields agents match on; if nothing fits, submit.html takes your request.
Can the agent do the whole sprint unattended?
Two gates stay human by design: the printer, and the FormLM login (or its headless fallback handoff).