Agent Skills: init-config

Initialization

Run phases in order because later phases depend on saved choices from earlier phases.

Phase 1: Environment Discovery

Goal: detect installed languages, tools, and package managers.

Actions:

1. Detect installed languages (Node.js, Python, Rust, Go, Java, Docker, and others).
2. Detect package manager options for selected ecosystems.
3. Store detection results for Phase 4 option generation.

Phase 2: Developer Profile

Goal: detect or collect developer identity for personalization.

Actions:

1. Run git config user.name and git config user.email to detect developer info.
2. If both are detected, store as developer_name and developer_email for renderer arguments and proceed to Phase 3.
3. If either is missing, ask the user directly through conversation:
   • If name is missing: "I couldn't detect your name from git config. What name would you like to use? (or reply 'skip' to omit)"
   • If email is missing: "I couldn't detect your email from git config. What email would you like to use? (or reply 'skip' to omit)"
4. Store the user's responses (or empty values if skipped) as developer_name and developer_email for renderer arguments.

Phase 3: Testing Methodology

Goal: choose the testing approach.

Actions: Ask with header Testing Methodology:

• BDD first, then TDD (Recommended) -> BDD-driven TDD with Red-Green-Refactor
• BDD only -> Gherkin scenarios without TDD cycle
• TDD only -> Test-driven development without BDD scenarios
• None -> No specific testing methodology Store the choice as testing_mode (bdd-tdd | bdd | tdd | none) for Phase 7 renderer arguments.

Phase 3.5: Memory Management (Optional)

Goal: decide whether to add CLAUDE.md memory instructions.

Actions: Ask with header Memory:

• Skip (Recommended)
• Include memory rules Store boolean include_memory for renderer arguments. Do not append memory text manually.

Phase 4: Technology Stack & Package Manager Selection

Goal: choose languages and package managers.

Actions:

1. Use AskUserQuestion (multiSelect: true) for technology stacks.
2. Generate options from detected technologies and mark detected ones as recommended.
3. For selected languages with multiple managers, ask preference:

• Node.js: npm, pnpm (Recommended), yarn, bun.
• Python: pip, uv (Recommended), poetry.
• Only show managers detected on the machine.

4. Store ordered stack selections as language:::package_manager for renderer arguments.

Phase 5: Renderer Input Preparation

Goal: prepare deterministic renderer inputs from user selections.

Actions:

1. Normalize each selected stack into ordered renderer input format language:::package_manager.

• Keep selection order unchanged.

2. For languages without explicit package manager choice, use language::: (empty manager).
3. Keep language keys exact (Node.js, Python, Rust, Swift, Go, Java) when available; do not invent aliases.
4. Never call online search in this phase.

Phase 6: Style Preference

Goal: choose emoji usage policy in generated CLAUDE.md.

Actions: Ask with header Style:

• No Emojis (Recommended)
• Use Emojis Store boolean use_emojis for renderer arguments that control emoji policy text in output.

Phase 7: Assembly & Generation

Goal: generate final content through one renderer script.

Actions:

1. Run ${CLAUDE_PLUGIN_ROOT}/scripts/render-claude-config.sh with:

• --target-file $HOME/.claude/CLAUDE.md
• --testing-mode <bdd-tdd|bdd|tdd|none> (from Phase 3)
• --include-memory <true|false>
• --use-emojis <true|false>
• Optional --developer-name and --developer-email
• Repeated --stack "language:::package_manager" entries from Phase 5

2. Let the renderer handle all assembly concerns: fragment assembly, testing content injection, developer profile, technology stack section, optional memory section, and final write.
3. Do not manually post-edit renderer output; if output shape is wrong, fix inputs or renderer behavior.

Phase 8: Write CLAUDE.md

Goal: report write and backup results from renderer.

Actions:

1. Use renderer output to confirm:

• Target path written.
• Backup path created when an existing target file was present.

2. Report:

• File and backup locations.
• Developer info and testing mode.
• Selected technology stacks and package managers.
• Renderer rule-application summary: which stacks received local rule lines and which did not.

Best Practices

• Keep workflow progressive and deterministic.
• Prefer local references over generated prose for stack guidance.
• Keep generated constraints concise and enforceable.
• Always back up existing files before overwrite.