AI_Workflow

Chapter 03 Exercise — Install the Starter Kit and Write Your First Custom Command

Time budget: 30 minutes Success criterion: All 30 starter-kit commands installed in a project of yours, AND you have written one new custom command that automates a prompt you’ve typed at least twice.

Part 1 — Install the starter kit (10 min)

1. Pick a project

Use tsd-exercise-01 from Chapter 01, or any real repo where you’ll use Claude Code:

cd ~/repos/tsd-exercise-01
mkdir -p .claude/commands

2. Copy all 30 starter-kit commands

cp ~/repos/AI_Workflow/03-slash-commands/starter-kit/.claude/commands/*.md \
   .claude/commands/

3. Verify they loaded

claude

Inside Claude, type / and look at the command list. You should see all 30 commands. Test the simplest one:

/prep

Expected: Read project context. Ready.

Then test a few from each family:

Test	Family	What you should see
`/test`	Testing	Lists test runner detected, runs the suite
`/code-review`	Quality	Starts a 5-phase audit of the working tree
`/status`	Session	Repo health summary

4. (Optional) Initialize SpecKit

If you plan to use the 10 SpecKit commands (/speckit.specify etc.), initialize SpecKit in this project:

# Outside Claude, in the same directory:
uvx --from git+https://github.com/github/spec-kit.git@v0.5.0 specify init

This creates .specify/ with the scripts and templates the SpecKit commands depend on. Skip this step if you only want the non-SpecKit families today.

5. Commit the starter kit

Back inside Claude:

/commit

You should see lint + type-check run, then a commit land with a message like chore: install 30-command starter kit.

Part 2 — (Optional) Install the full 65-command kit (5 min)

If you want the orchestration, governance, wireframe pipeline, and knowledge curation commands (Chapter 05 territory), run the install script:

bash ~/repos/AI_Workflow/03-slash-commands/install-full-kit.sh

This copies the remaining 35 commands from ~/repos/Claude_Commandz/global/.claude/commands/ into your global ~/.claude/commands/ (not per-project — these are personal commands).

You only do this once, ever. After this you have all 65 of TurtleWolfe’s commands available globally.

Don’t run this on Day 1 unless you know you need it. Most of these commands need infrastructure that doesn’t exist in your project yet (the 27-terminal tmux session, the wireframe validator script, etc.). They’ll show up in /help but won’t do anything useful until you set up Chapter 05.

Part 3 — Write your own command (15 min)

1. Identify a repetitive prompt

Think back to the last 5-10 sessions with Claude Code. What’s a prompt you’ve typed more than once? Candidates:

“Show me every TODO comment in the codebase”
“List all files larger than 500 lines”
“Find every component without a test file”
“Show me the last 10 failing test runs”
“Scaffold a new component with the 5-file pattern”
“Run docker compose down && docker compose up --build”
“Generate the weekly standup summary from git log”

Pick one. If you don’t have a repetitive prompt yet (Day 1!), pick “Show me every TODO comment” — that’s a good starter.

2. Sketch the command

Create .claude/commands/<your-command-name>.md:

$EDITOR .claude/commands/find-todos.md

Use this skeleton:

---
description: One-sentence description shown in the command list
scope: personal
---

Short prose explaining what this command does and why.

## Instructions

1. Step one — specific action
2. Step two — specific action
3. Step three — specific action

## Output

What Claude should report when done. Be specific.

## Notes

- Any edge cases
- When NOT to use this command

3. Write the command body

For the “find TODOs” example:

---
description: List every TODO, FIXME, HACK, and XXX comment in the codebase with file:line
scope: personal
---

Find all marker comments and report them in a browsable list.

## Instructions

1. Search the repo (excluding node_modules, .next, dist, build) for comments containing:
   - TODO
   - FIXME
   - HACK
   - XXX
   - BUG

2. For each match, capture:
   - File path
   - Line number
   - The comment text

3. Group matches by file and sort by file path.

4. For each file, list the matches indented under it with the file:line format so the user can click to jump.

## Output

Format:

\`\`\`
src/lib/auth.ts
  src/lib/auth.ts:42 — TODO: handle refresh token expiry
  src/lib/auth.ts:87 — FIXME: race condition on concurrent logout

src/components/Button/Button.tsx
  src/components/Button/Button.tsx:15 — HACK: inline styles until theme lands
\`\`\`

End with a count: `Found N markers in M files.`

If no matches, report `No marker comments found.`

## Notes

- Exclude third-party code (node_modules, vendor/, etc.)
- Do not fix the TODOs — this is a report-only command

4. Test your command

Back in Claude:

/find-todos

(or whatever you named it)

Claude should follow the instructions and produce the output format you specified.

If it doesn’t work exactly right, the problem is almost always vague instructions. Sharpen them and try again. This is the core skill: writing commands specific enough that Claude behaves the same every time.

5. Commit your new command

/commit

You should see a commit like feat: add /find-todos command.

Rubric

All 30 starter-kit commands live in your project’s .claude/commands/
/prep, /commit, /test all work without errors
You wrote a new .md file in .claude/commands/ with frontmatter (description, scope)
Your new command has an ## Instructions section with numbered steps
Your new command has an ## Output section describing the expected result
Your new command runs successfully and produces output matching what you specified
Both the starter kit AND your new command are committed

If any box is unchecked, the exercise isn’t done.

What you learned

Slash commands are just markdown files. There’s no framework, no API, no compiler. Write a file, use it.
Specificity is everything. “Find TODOs” is vague. “Search excluding node_modules, group by file, format as file:line” is a command.
The starter kit is the starter kit because you’ll use it daily. Your custom commands grow from pain — every command you write solves a repeated annoyance.
The full 65-command kit is one script away. Don’t install it until you need it.

Next: Chapter 04 — The SpecKit Loop

This site is open source. Improve this page.