Claude Code Starter

Generate a complete .claude/ configuration for any project by analyzing its tech stack.

Overview

This skill performs 6 steps:

1. •Detect the project's tech stack (languages, frameworks, tools)
2. •Check for existing .claude/ configuration
3. •Generate settings.json with stack-specific permissions
4. •Generate skills, agents, rules, and commands from templates
5. •Analyze the codebase deeply and write CLAUDE.md
6. •Summarize all files created

Step 1: Detect Tech Stack

Analyze the current project root to detect languages, frameworks, and tools.

Languages

Check for these files and extensions at the project root:

The first detected language is the primary language.

Frameworks

JS/TS (from package.json dependencies/devDependencies, order matters — first match wins for frontend):

• •Frontend (mutually exclusive, check in order): next -> Next.js, nuxt/nuxt3 -> Nuxt, @sveltejs/kit -> SvelteKit, svelte -> Svelte, astro -> Astro, @remix-run/react -> Remix, gatsby -> Gatsby, solid-js -> Solid, @angular/core -> Angular, vue -> Vue, react -> React
• •Backend (additive): @nestjs/core -> NestJS, express -> Express, fastify -> Fastify, hono -> Hono, elysia -> Elysia, koa -> Koa
• •CSS/UI (additive): tailwindcss -> Tailwind, components.json file -> shadcn, @chakra-ui/react -> Chakra, @mui/material -> MUI
• •Database/ORM (additive): prisma/@prisma/client -> Prisma, drizzle-orm -> Drizzle, typeorm -> TypeORM, sequelize -> Sequelize, mongoose -> Mongoose

Python (from requirements.txt or pyproject.toml): fastapi, django, flask, starlette, sqlalchemy

Ruby (from Gemfile): rails, sinatra

Go (from go.mod): gin-gonic -> Gin, labstack/echo -> Echo, gofiber -> Fiber

Rust (from Cargo.toml): actix, axum, rocket

Swift/iOS: Check Package.swift for vapor. Check .xcodeproj/.xcworkspace + Swift files for SwiftUI indicators (ContentView, App.swift, @main struct) and UIKit indicators (ViewController, AppDelegate, SceneDelegate). Scan first 10 Swift files for import SwiftData/@Model and import Combine/PassthroughSubject.

Android/Kotlin (from build.gradle/build.gradle.kts including app/ variants): Check for com.android/android {, then compose/androidx.compose -> Jetpack Compose (else Android Views), room/androidx.room -> Room, hilt/dagger.hilt -> Hilt, ktor -> Ktor. Also check for spring -> Spring, quarkus -> Quarkus.

Tools

Package Manager (check lock files in order): bun.lockb/bun.lock -> bun, pnpm-lock.yaml -> pnpm, yarn.lock -> yarn, package-lock.json -> npm, poetry.lock -> poetry, Pipfile.lock/requirements.txt -> pip, Cargo.lock -> cargo, go.sum -> go, Gemfile.lock -> bundler, pom.xml -> maven, build.gradle/.kts -> gradle

Testing (from dependencies/config files): vitest, jest, mocha, @playwright/test, cypress, bun test in scripts -> bun-test, pytest.ini/conftest.py -> pytest, go.mod -> go-test, Cargo.toml -> rust-test, Gemfile -> rspec

Linter (check config files first, then dependencies): eslint.config.*/.eslintrc* -> eslint, biome.json/biome.jsonc -> biome, ruff.toml/.ruff.toml -> ruff, .flake8/setup.cfg -> flake8

Formatter: .prettierrc*/prettier.config.* -> prettier, biome.json/biome.jsonc -> biome, pyproject.toml -> black

Bundler: vite.config.* -> vite, webpack.config.* -> webpack, tsup.config.* -> tsup, rollup.config.* -> rollup, esbuild* -> esbuild; from deps: parcel, @vercel/turbopack -> turbopack, @rspack/core -> rspack

CI/CD: .github/workflows/ dir -> github-actions, .gitlab-ci.yml -> gitlab-ci, .circleci/ dir -> circleci, .travis.yml -> travis, azure-pipelines.yml -> azure-devops, Jenkinsfile -> jenkins

Docker: Dockerfile, docker-compose.yml, docker-compose.yaml

Monorepo: pnpm-workspace.yaml, lerna.json, nx.json, turbo.json, workspaces in package.json, packages/ or apps/ directories

Step 2: Check Existing Configuration

1. •Check if .claude/ directory exists
2. •If it exists, ask the user: Update (preserve existing task state) or Overwrite (replace all)?
3. •If .claude/state/task.md exists, always preserve it regardless of user choice

Step 3: Generate Settings

Create .claude/settings.json using the permission patterns from settings-patterns.md.

1. •Start with base permissions (always included)
2. •Add language-specific permissions for each detected language
3. •Add testing framework permissions
4. •Add linter and formatter permissions
5. •Add Docker permissions if Docker detected
6. •Deduplicate the final array

Step 4: Generate Skills, Agents, Rules, Commands

Use the templates from the reference files to generate all applicable files.

Skills

See templates-skills.md for all skill templates.

Always generate (8 core skills):

• •.claude/skills/pattern-discovery.md
• •.claude/skills/systematic-debugging.md
• •.claude/skills/testing-methodology.md (dynamic: uses detected testing framework)
• •.claude/skills/iterative-development.md (dynamic: uses detected test/lint commands)
• •.claude/skills/commit-hygiene.md
• •.claude/skills/code-deduplication.md
• •.claude/skills/simplicity-rules.md
• •.claude/skills/security.md (dynamic: JS/TS and Python sections are conditional)

Conditionally generate (framework skills):

• •.claude/skills/nextjs-patterns.md (if Next.js detected)
• •.claude/skills/react-components.md (if React detected AND Next.js NOT detected)
• •.claude/skills/fastapi-patterns.md (if FastAPI detected)
• •.claude/skills/nestjs-patterns.md (if NestJS detected)
• •.claude/skills/swiftui-patterns.md (if SwiftUI detected)
• •.claude/skills/uikit-patterns.md (if UIKit detected)
• •.claude/skills/vapor-patterns.md (if Vapor detected)
• •.claude/skills/compose-patterns.md (if Jetpack Compose detected)
• •.claude/skills/android-views-patterns.md (if Android Views detected)

Agents

See templates-agents.md for agent templates.

Always generate (2 agents):

• •.claude/agents/code-reviewer.md (dynamic: lint command based on detected linter)
• •.claude/agents/test-writer.md (dynamic: test command based on detected framework)

Rules

See templates-rules.md for rule templates.

• •.claude/rules/typescript.md (if TypeScript detected)
• •.claude/rules/python.md (if Python detected)
• •.claude/rules/code-style.md (always — references detected formatter/linter)

Commands

See templates-commands.md for command templates.

Always generate (5 commands):

• •.claude/commands/task.md
• •.claude/commands/status.md
• •.claude/commands/done.md
• •.claude/commands/analyze.md
• •.claude/commands/code-review.md

State

• •Create .claude/state/task.md with empty initial state (only if it does not already exist)

Step 5: Deep Codebase Analysis

Perform the full codebase analysis described in analysis-prompt.md.

This is the most critical step. Follow ALL phases:

1. •Phase 1: Discovery — Read actual project files to gather specific information
2. •Phase 2: Generate — Write .claude/CLAUDE.md using ONLY discovered information
3. •Phase 3: Verify — Check quality before finalizing

The CLAUDE.md must include the detected tech stack summary table and reference the generated skills, agents, rules, and commands in its structure. Include actual commands from package.json scripts (or equivalent), real file paths, and observed code conventions.

The CLAUDE.md MUST reference:

• •The project name and description (from package.json/README/equivalent)
• •The detected tech stack with specific configuration notes
• •Actual directory structure (depth 3)
• •Real commands from scripts
• •Code conventions observed from reading source files
• •The generated skills, agents, and commands

Step 6: Summary

Output a summary to the user:

## Claude Code Starter - Setup Complete

### Tech Stack Detected
- **Language**: {primary language}
- **Framework**: {primary framework}
- **Package Manager**: {package manager}
- **Testing**: {testing framework}
- **Linter**: {linter}
- **Formatter**: {formatter}

### Files Created
{list all files created/updated with paths}

### Framework-Specific Skills
{list any conditional skills that were generated}

### Next Steps
1. Review `.claude/CLAUDE.md` and adjust any inaccuracies
2. Review `.claude/settings.json` permissions
3. Try `/task <description>` to start your first task
4. Use `/status` to check progress
5. Use `/done` to mark tasks complete

Dynamic Value Reference

Test Commands

Lint Commands

State File Initial Content

For existing projects (source files detected):

# Current Task

## Status: Ready

No active task. Start one with `/task <description>`.

## Project Summary

{project name}{if description: - description}

**Tech Stack:** {summarized tech stack}

## Quick Commands

- `/task` - Start working on something
- `/status` - See current state
- `/analyze` - Deep dive into code
- `/done` - Mark task complete

For new/empty projects:

# Current Task

## Status: In Progress

**Task:** {description or "Explore and set up project"}

## Context

New project - setting up from scratch.

{if framework: **Framework:** formatted name}
{if language: **Language:** formatted name}

## Next Steps

1. Define project structure
2. Set up development environment
3. Start implementation

## Decisions

(None yet - starting fresh)