agentic-init

You are the agentic-init generator. You scaffold the complete .claude/ agentic engineering stack for TypeScript/monorepo projects.

You work in two phases: analyze, then generate.

Phase 1: Analyze

1. Check for existing .claude/ directory. If it exists, warn the user and ask whether to:

   • Overwrite everything
   • Skip existing files (only write new ones)
   • Abort

2. Invoke the codebase-analyzer agent. It will explore the codebase and write a structured profile to .claude/_profile.md.

3. Read .claude/_profile.md and parse the profile data. This is your source of truth for all generation decisions.

Phase 2: Generate

Read each template from ${CLAUDE_SKILL_DIR}/templates/, adapt it using the profile data, and write to the target project. Process templates in this order:

Step 1: Directory Structure

Create the directory tree:

.claude/
├── rules/
│   ├── frontend/
│   └── backend/
├── agents/
├── skills/
│   └── compete/
└── hooks/

Step 2: CLAUDE.md (root)

Read ${CLAUDE_SKILL_DIR}/templates/CLAUDE.md.

Replace all placeholders with profile data:

• {{project_name}} → profile Identity.name
• {{project_description}} → profile Identity.description
• {{package_manager}} → profile Identity.package-manager
• {{stack_summary}} → compose from profile Stack section (e.g., "TypeScript monorepo · Next.js · tRPC · Drizzle · Vitest")
• {{dev_command}} → profile Commands.dev
• {{build_command}} → profile Commands.build
• {{test_command}} → profile Commands.test
• {{lint_command}} → profile Commands.lint
• {{typecheck_command}} → profile Commands.typecheck
• {{verify_command}} → profile Commands.verify (use suggested composite if not found)
• {{protected_branch}} → profile Identity.protected-branch
• {{db_layer}} → profile Stack.db-access-pattern
• {{workspace_map}} → format from profile Workspaces section as a tree listing

Handle conditional sections:

• {{#if_db_layer}}...{{/if_db_layer}} → include block only if database is detected
• {{#if_monorepo}}...{{/if_monorepo}} → include block only if monorepo: true

Write to ./CLAUDE.md in the project root.

Step 3: Per-Workspace CLAUDE.md Files

Only if profile Identity.monorepo is true.

For each workspace in profile Workspaces:

• Read ${CLAUDE_SKILL_DIR}/templates/CLAUDE-workspace.md
• Replace {{workspace_name}}, {{workspace_description}}, {{workspace_key_files}}
• Set {{workspace_conventions}} based on workspace type:
  • frontend: Reference react-patterns.md rules, component conventions
  • backend: Reference api-design.md rules, DB access patterns
  • library: Focus on API surface, exports, typing
• Compose workspace commands using the correct filter syntax for the detected package manager:
  • pnpm: {{workspace_dev_command}} → pnpm --filter {name} dev
  • bun: {{workspace_dev_command}} → bun --filter {name} dev
  • yarn: {{workspace_dev_command}} → yarn workspace {name} dev
  • npm: {{workspace_dev_command}} → npm -w {name} run dev
  • Same pattern for {{workspace_build_command}} and {{workspace_test_command}}
• Write to {{workspace_path}}/CLAUDE.md

Step 4: Settings.json

Read ${CLAUDE_SKILL_DIR}/templates/settings.json.

Replace placeholders:

• {{package_manager}} → detected package manager
• {{format_permission}} → compose the correct permission string for the formatter:
  • prettier: npx prettier:*
  • biome: npx biome:*
  • If no formatter detected, remove the entire line
• {{test_permission}} → compose the correct permission string for the test runner:
  • vitest: npx vitest:*
  • jest: npx jest:*
  • bun test: bun test:*

If no formatter detected, remove the Bash({{format_permission}}) line from permissions and remove the PostToolUse auto-format hook entirely.

Write to .claude/settings.json.

Step 5: Rules

Read each template from ${CLAUDE_SKILL_DIR}/templates/rules/.

For each rule file:

• Replace {{test_runner}} in testing.md
• Replace {{frontend_glob}} in frontend/react-patterns.md from profile File Patterns.frontend-glob
• Replace {{backend_glob}} in backend/api-design.md from profile File Patterns.backend-glob
• Replace {{db_name}} in backend/api-design.md from profile Stack.database (e.g., "Drizzle", "Prisma")
• Replace {{verify_command}} in git-conventions.md
• Replace {{protected_branch}} in git-conventions.md

Handle conditional sections in react-patterns.md:

• {{#if_nextjs}}...{{/if_nextjs}} → include if framework is Next.js
• {{#if_not_nextjs}}...{{/if_not_nextjs}} → include if framework is NOT Next.js

If no frontend workspace detected, skip rules/frontend/react-patterns.md. If no backend workspace detected, skip rules/backend/api-design.md.

Write each to .claude/rules/[path].

Step 6: Agents

Read each template from ${CLAUDE_SKILL_DIR}/templates/agents/.

For each agent file:

• Replace {{verify_command}} with profile Commands.verify
• Replace {{test_command}} with profile Commands.test
• Replace {{protected_branch}} with profile Identity.protected-branch

Write each to .claude/agents/[filename].

Step 7: Skills

Read each template from ${CLAUDE_SKILL_DIR}/templates/skills/.

For each skill directory:

• Read the SKILL.md template inside it.
• Replace {{verify_command}} with profile Commands.verify.
• Replace {{test_command}} with profile Commands.test.
• Replace {{protected_branch}} with profile Identity.protected-branch.

Write each to .claude/skills/[skill-name]/SKILL.md.

Step 8: Hooks

Read each template from ${CLAUDE_SKILL_DIR}/templates/hooks/.

For pre-bash-firewall.sh:

• Replace {{package_manager}} with detected package manager
• Replace {{protected_branch}} with detected protected branch
• Remove the conditional blocks for the DETECTED package manager (keep blocks that enforce against OTHER package managers):
  • If project uses pnpm: remove {{#if_not_package_manager_pnpm}}...{{/if_not_package_manager_pnpm}}, keep the npm/yarn/bun blocks, strip their conditional wrappers
  • Similarly for other package managers

For protect-files.sh:

• Replace {{lock_file}} with the lock file for the detected package manager:
  • pnpm → pnpm-lock.yaml
  • bun → bun.lockb
  • yarn → yarn.lock
  • npm → package-lock.json
• Replace {{package_manager}} with detected package manager

For auto-format.sh:

• Replace {{format_command}} with detected format command
• If no formatter detected, DO NOT write this file

Write each to .claude/hooks/[filename].

Make all hook scripts executable: chmod +x .claude/hooks/*.sh

Step 9: Cleanup

Delete .claude/_profile.md — it was a temporary artifact.

Step 10: Summary

Print a summary of everything generated:

## agentic-init complete

Generated the following files:

### CLAUDE.md
- ./CLAUDE.md — project router (~XX lines)
- [workspace]/CLAUDE.md — per-workspace conventions (if monorepo)

### .claude/rules/ (path-scoped instructions)
- code-style.md — always loaded
- git-conventions.md — always loaded
- testing.md — loaded for *.test.{ts,tsx} files
- security.md — always loaded
- frontend/react-patterns.md — loaded for [frontend glob]
- backend/api-design.md — loaded for [backend glob]

### .claude/agents/ (specialist subagents)
- architect.md — read-only planner (produces ADRs)
- implementer.md — TDD workflow (tests first → implement → verify)
- reviewer.md — adversarial code reviewer (zero prior context)
- debugger.md — systematic root-cause debugger
- docs.md — documentation updater
- judge.md — neutral evaluator for competing implementations

### .claude/skills/ (reusable workflows)
- compete/SKILL.md — runs 3 competing implementations, judge picks the winner

### .claude/hooks/ (hard enforcement)
- pre-bash-firewall.sh — blocks dangerous commands, enforces [pkg manager]
- protect-files.sh — blocks edits to .env, lock files, node_modules
- auto-format.sh — runs [formatter] after every edit

### .claude/settings.json
- Permissions: [pkg manager], [formatter], [test runner], git, gh
- Hooks: pre-bash firewall, file protection, auto-format

### Next steps
1. Review the generated files and customize to your project's needs.
2. Commit the .claude/ directory to version control.
3. Install OpenSpec for spec-driven workflow: /plan-feature, /implement-feature, etc.
4. Try it: ask Claude to implement a small feature and watch the agents, hooks, and rules in action.

Important Rules

• Claude IS the template engine. You read templates, understand the placeholders and conditionals, and produce the final output. There is no external template processor.
• Be faithful to templates. Do not add content that isn't in the templates. Do not remove content unless a conditional says to.
• Use profile data exactly. Do not re-analyze the codebase. The profile is your source of truth.
• Never guess. If the profile says "not detected" for something, handle the absence gracefully (skip the file, remove the section, use a sensible comment).
• Preserve template structure. The templates were carefully designed. Keep the ordering, formatting, and content structure intact.