Skip to content
Claude with Skills

Exercise 01: Build a Custom Deploy Skill

Build a deploy skill tailored to a project of your choice. This exercise puts together the fields covered in lessons 02, 03, 04, and 05 into a single working skill.

Requirements

Your skill must:

  1. Set disable-model-invocation: true — deploying should never happen automatically.
  2. Include an argument-hint showing what argument the user should provide (typically an environment name: staging, production, etc.).
  3. Set allowed-tools with only the commands your deploy workflow actually needs. Do not use Bash(*).
  4. Use at least one !`cmd` injection for a pre-flight check that runs before Claude does anything else.

Choose your stack

Pick one of these or adapt to your own project:

Node.js / npm

Pre-flight: git status, git log --oneline -3 Build: npm run build Test: npm test Publish: git tag + git push

Python / pip

Pre-flight: git status, git log --oneline -3 Test: pytest Build: python -m build Publish: git tag + git push

Docker

Pre-flight: git status, docker info Build: docker build Push: docker push Deploy: docker compose up -d (or your orchestration tool)

Static site (Netlify, Vercel, etc.)

Pre-flight: git status, git log --oneline -3 Build: npm run build Deploy: netlify deploy --prod or vercel --prod

Validation

After writing the skill, install it in your personal skills directory:

Terminal window
mkdir -p ~/.claude/skills/my-deploy
cp SKILL.md ~/.claude/skills/my-deploy/

Then open Claude Code in a project and test these two scenarios:

Scenario A: Uncommitted changes

  1. Make an edit to any file without committing.
  2. Invoke the skill: /my-deploy staging
  3. Confirm the skill detects the uncommitted changes from the !`git status` output and aborts with a clear error message. It should not proceed to run tests or build.

Scenario B: Clean working tree

  1. Commit all pending changes.
  2. Invoke the skill: /my-deploy staging
  3. Confirm the skill runs the full workflow in order.

Checklist

  • disable-model-invocation: true is set
  • argument-hint is present
  • allowed-tools lists specific commands, not Bash(*)
  • At least one !`cmd` injection is in the body
  • Scenario A: skill aborts on uncommitted changes
  • Scenario B: skill runs the full workflow on a clean tree

Solution

A worked solution for a Python project is in solutions/01-deploy-skill/SKILL.md. Read the exercise before looking at the solution.