Lab 3 - Skills

Install ready-made Robot Framework skills, then write one of your own that enforces a convention of this repository, and check that it loads when it should, and only then.

Module	3 - Agent Skills
Time	25 minutes
Shop preset	clean
You need	Lab 0 done. Your AGENTS.md from Lab 2 helps, but this lab works without it
You start from	main, plus your work from Lab 2 if you have it

A skill is a folder with a SKILL.md. The agent reads only its description until a task matches it; then it loads the rest. Rule of thumb: what is always relevant goes into AGENTS.md; what is relevant for one kind of task becomes a skill.

Steps

1. Install the Robot Framework Agent Skills into this repository. The version is pinned in SETUP.md:

   Claude Code	Codex	GitHub Copilot
   uvx rf-agentskills@0.6.0 install --agent claude-code --scope project --project . --what skills	same, with --agent codex	same, with --agent copilot
   lands in .claude/skills/	lands in .agents/skills/	lands in .claude/skills/, which Copilot also reads

2. See what you got. Start a new agent session and ask:

   Which skills do you have for Robot Framework, and what is each one for? One line each.

3. Rerun Lab 2's prompt, now with the skills installed:

   Write a Robot Framework test for criterion WEB-002_AC-5 of this repository and save it as results/lab-03/with-skills.robot. Do not run it, and do not change any other file.

   Compare it with results/lab-02/after.robot, if you have it. Did a skill show up in the session? Did the keywords it chose change?

4. Pick a convention from docs/conventions.md for your own skill. A good first choice is convention 2, Locators live in resources. Any other works too.

5. Copy the template into your agent's skill folder, under a name for your convention, for example locators-in-resources:

   Claude Code	Codex	GitHub Copilot
   cp -r skills/template .claude/skills/locators-in-resources	cp -r skills/template .agents/skills/locators-in-resources	cp -r skills/template .github/skills/locators-in-resources

   The template is a working skill for convention 1. Read it once as it is.

6. Make it yours. In the copied SKILL.md:

   • set name to the folder's name;
   • rewrite description: what the skill checks, and when to use it, in the words a prompt would contain;
   • replace the rule, the reason and the examples with your convention's;
   • delete the template's comment.

   Leave the script in scripts/ alone for now; it belongs to the stretch goal.

7. Check that it triggers. New session, then a prompt that should load it:

   Review tests/ui/catalogue.robot against the conventions of this repository.

   The session shows that your skill loaded:

   Claude Code	Codex	GitHub Copilot
   prints Skill(<name>)	reads your skill's SKILL.md	names the skill it invokes

   Did the review find anything?

8. Check that it stays quiet. A prompt about something else:

   What does uv run --no-sync python -m shop reset do? Answer in two sentences.

   Your skill must not load. If it does, your description is too broad: make it more specific and repeat steps 7 and 8.

Stretch

Give your skill teeth. Change scripts/check.py in your skill's folder so that it checks your convention instead of convention 1. For convention 2, for example, report every argument in a test file that looks like a locator. Then tell the agent in SKILL.md to run it, and repeat step 7.

If you let your agent make the change: the skill folders are agent configuration, and agents guard them. Claude Code asks before it writes into .claude/, and Codex's sandbox keeps .agents/ read-only until you approve the write. Approve it, or edit the files yourself.

Compare with the reference

When you are done, compare your result with the reference: what the rehearsal produced, why it is a good result, and the answers the debrief covers. Open it after the lab: it gives the answers away.

If your agent fails

Steps 4 to 6 need only an editor. For the rest, follow the recorded walkthrough of this lab.