skill tutorial · agent
Workshop skills in GitHub Copilot
Copilot's agent mode reads instruction files — .github/copilot-instructions.md repo-wide, .github/instructions/*.instructions.md scoped by apply-mode. A WorkshopSkill bundle plus one of those files is a working skills drawer; the repo becomes the loader.
Add the folder and the instruction
repo shape
training-drawer/ skills/the-orange/ (from the-orange.zip) .github/ copilot-instructions.md curl -LO https://workshopskill.com/en/activities/the-orange/files/the-orange.zip unzip the-orange.zip && mkdir -p skills && mv the-orange skills/
copilot-instructions.md
the repo instruction file
## Workshop prep Skills live in skills/<name>/SKILL.md (Agent Skills format, index: https://workshopskill.com/en/skills.json). When asked to prep an activity: match the brief to metadata (category, group_size, duration_minutes, setting), open only that SKILL.md, run its curl steps as printed, treat the assessment JSON as the measure spec, ask before formlm-cli.
First run
Agent mode, inside the repo: prep The Orange for two teams of three, conflict module, 20 minutes
. Copilot opens the skill, pulls the hidden briefs (flesh vs peel) and the debrief, and states the integration lesson from the folder — then asks before touching the live measure, where a Copilot CLI session with a shell can complete the formlm-cli steps.
Notes
- Code-scoped instructions (
applyTo) are for file patterns; for behavior like prep-matching the repo-wide file is the right scope. - Copilot CLI runs the same pattern with a live shell — the curl steps and the optional formlm-cli steps execute as printed.
- Private repos: bundles are CC BY 4.0; committing them is permitted and keeps paper and instructions in one review.
keep going
questions
Does Copilot support agent skills natively?
The Agent Skills convention is vendor-neutral and adoption is spreading; where Copilot has no loader, the instruction-file pattern above is the faithful route — the SKILL.md is unchanged.
Instruction budget?
Repo-wide instructions load per session — keep the file short and pointer-style; skills load on demand.
Can Copilot print?
Humans print, agents parse — Copilot stages the PDFs; the printer is yours.