My Claude Code workshop from idea to release

1. Step 1 Opus

   Idea

   Should we build at all? The agent looks at what Google, published tools and my own projects already do, compares the possible forms and names what could sink the idea. It may conclude that nothing should be written.

2. Step 2 Opus

   Scoping

   Who for, what problem, which data and how much, which Google limits. It sets the acceptance criteria the tests will check.

3. Step 3 Opus

   Design

   Files, tabs and columns, resuming after 6 minutes, plan then apply, OAuth scopes. It reuses my shared base instead of reinventing it.

4. Step 4 Sonnet

   Development

   It writes the code and its test cases, then fixes what the next steps send back. It releases nothing.

5. Step 5 Opus

   Testing

   A fresh agent, who did not write the code, tests the tool outside Google on data shaped exactly like the API output. At most three rounds with development.

6. Step 6 Opus

   Review

   It looks for what the code allows too much: execution identity, injection into HTML pages, overly broad scopes, secrets.

7. Step 7 Opus

   Documentation

   It discovers the tool as a newcomer would and writes the guide. It checks that every menu and message it quotes exists word for word.

8. Step 8 Sonnet

   Release

   First a plan: commits, changelog, version number, what will leave the machine. Only then, with my approval, does it release.

Testing, review and documentation go to agents that never saw the developer’s reasoning. Whoever wrote the code shares its blind spots: they test what they meant to do, and document what they already know.

1. 1. Write your rules

   A CLAUDE.md file at the project root, or in ~/.claude for every project. Claude reads it at the start of each session. Only put in what has already cost you once: one rule per mistake paid for.

   CLAUDE.md

   # Règles de mes projets Apps Script
   - Jamais de clasp push direct : d'abord la vérification.
   - Toute date lue dans Sheets passe par normaliserDate_().
   - Toute correction s'accompagne du test qui l'aurait attrapée.

2. 2. One agent per role

   A Markdown file in .claude/agents/. The header sets its name, when to use it, its model and the tools it may not use; the body describes its job. Here the testing agent cannot edit the code: it can only report.

   .claude/agents/essais.md

   ---
   name: essais
   description: Éprouve l'outil contre ses critères d'acceptation, sans avoir écrit le code.
   model: opus
   disallowedTools: Edit
   ---
   Tu n'as pas écrit ce code. Lance le banc d'essai, confronte chaque
   critère au résultat, et rends ACCEPTÉ ou REFUSÉ avec les défauts.

3. 3. A check that blocks

   A hook runs your script before each command. The script receives the command, and if it exits with code 2, Claude Code refuses it and shows the message to Claude, which has to fix things. That is the difference between an instruction and a guard.

   .claude/settings.json

   {
     "hooks": {
       "PreToolUse": [{
         "matcher": "Bash",
         "hooks": [{ "type": "command", "command": "node .claude/garde.js" }]
       }]
     }
   }
   
   // .claude/garde.js : le code de sortie 2 refuse la commande
   const { tool_input } = JSON.parse(require('fs').readFileSync(0, 'utf8'));
   if (/clasp (push|deploy)/.test(tool_input.command ?? '')) {
     console.error('Envoi bloqué : lancez d’abord la vérification.');
     process.exit(2);
   }

4. 4. The chain in one command

   A skill in .claude/skills/chaine/: Claude runs it when you type /chaine. It chains the agents and holds the gates, since agents cannot talk to you directly.

   .claude/skills/chaine/SKILL.md

   ---
   name: chaine
   description: Mène un outil de l'idée à la publication, agent par agent.
   ---
   Lance les agents dans l'ordre : idee, cadrage, conception,
   developpement, essais, relecture, documentation, publication.
   Après chacun, montre-moi son verdict et attends mon accord.
   Si les essais refusent, renvoie les défauts au développement.

5. 5. Bundle it

   Once the framework works, a plugin makes it reusable: the same agents, skills and hooks in every new project, or for a whole team, without copying files around.

   plugin

   mon-atelier/
   ├── .claude-plugin/plugin.json
   ├── agents/        essais.md, relecture.md…
   ├── skills/        chaine/SKILL.md
   └── hooks/         hooks.json

Start small: rules and a hook are enough to avoid the mistakes you already know. Agents and the chain come when a project calls for them. The exact names and fields are described in the Claude Code documentation.