paxify

agentic-init

Claude Code plugin that scaffolds the full .claude/ agentic engineering stack for TypeScript/monorepo projects.

Defense-in-depth: CLAUDE.md, path-scoped rules, specialist agents, lifecycle hooks, and scoped permissions.

Install

/plugin marketplace add pax-k/paxify-plugins
/plugin install agentic-init@paxify

Usage

/paxify:agentic-init

The plugin analyzes your codebase and generates a complete .claude/ directory tailored to your project.

What It Generates

CLAUDE.md

Project constitution (~80-100 lines). Commands, workflow, key rules. Acts as a router to detailed context — not a novel.

.claude/rules/

Path-scoped instructions that load only when Claude touches matching files:

• code-style.md — TypeScript strict, naming, imports (always loaded)
• git-conventions.md — conventional commits, PR workflow (always loaded)
• testing.md — test runner, assertion rules (loaded for test files)
• security.md — OWASP, input validation, no secrets (always loaded)
• frontend/react-patterns.md — framework-specific patterns (loaded for frontend files)
• backend/api-design.md — API design, DB access layer (loaded for backend files)

.claude/agents/

Specialist subagents with isolated context and scoped tools:

• architect.md — read-only planner, produces ADRs
• implementer.md — TDD workflow: tests first → implement → verify
• reviewer.md — adversarial reviewer with zero prior context
• debugger.md — systematic root-cause debugging
• docs.md — documentation updater
• judge.md — neutral evaluator for competing implementations

.claude/skills/

Reusable workflow templates:

• compete/SKILL.md — runs 3 competing implementations (minimal, thorough, creative), judge picks the winner

.claude/hooks/

Shell scripts at lifecycle events. Exit 2 = blocked, no negotiation:

• pre-bash-firewall.sh — blocks dangerous commands, enforces package manager
• protect-files.sh — blocks edits to .env, lock files, node_modules
• auto-format.sh — runs formatter after every edit

.claude/settings.json

Permissions allowlist + hook wiring. Generous permissions for safe operations, hooks catch the dangerous ones.

How It Works

Phase 1: Analyze — A read-only codebase-analyzer agent explores your project and produces a structured profile: package manager, monorepo structure, framework, test runner, formatter, linter, database layer, git conventions, file patterns.

Phase 2: Generate — The /agentic-init skill reads the profile and adapts template files to your project. Every placeholder is replaced with detected values. Conditionals handle monorepo vs single-project, Next.js vs other frameworks, etc.

What It Detects

Category	Detects
Package manager	pnpm, bun, yarn, npm (from lock files)
Monorepo	Workspaces, workspace types (frontend/backend/library)
Framework	Next.js, Remix, Astro, Express, Fastify, Hono, tRPC
Test runner	Vitest, Jest, Bun test, Playwright, Cypress
Formatter	Prettier, Biome
Linter	ESLint, Biome, oxlint
Database	Drizzle, Prisma, TypeORM, Knex
Git	Commit style, protected branches, CI config

Workflow Integration

This plugin generates the guardrails layer. For spec-driven workflow (plan → implement → review), use OpenSpec.

Customization

After generation, review and customize:

1. Edit CLAUDE.md to add project-specific rules
2. Adjust path globs in rules to match your directory structure
3. Add/remove agent capabilities as needed
4. Tune hook scripts for your team's conventions
5. Modify settings.json permissions for your tooling

License

MIT