bp-context

name

bp-context

description

Scan project and generate/audit CLAUDE.md files with stack auto-detection and Context7 docs. Triggers on "/bp-context", "/context", "generate context", "audit CLAUDE.md", "scan project", or any request to create or update CLAUDE.md documentation files. Also triggers on "brownfield", "onboard project", "onboard this project", "onboard existing project", "scan existing project", "update context docs", "context scan", "check CLAUDE.md", "stale documentation", or "refresh project docs". Runs parallel workers per directory. Also audits README.md for staleness.

Context: Generate & Audit CLAUDE.md Files

Language

Read blueprint/.config.yml → language. If auto, detect from the user's messages. All generated content MUST be in the detected language. Skill instructions stay in English — only output changes.

Scan the project, auto-detect the tech stack, query framework documentation via Context7, and generate a tree of lean, focused CLAUDE.md files. Runs multiple subagents in parallel — one per directory cluster.

When to Run

• After /start on an existing codebase (brownfield onboarding)
• When onboarding to a project for the first time
• After adding major new packages or refactoring a directory
• When a CLAUDE.md feels stale or has incorrect info
• When the user says "brownfield", "onboard project", or "scan existing project"

Critical Rules

• NEVER overwrite existing CLAUDE.md without --force — but ALWAYS audit for staleness
• NEVER run full test suite — targeted tests only
• Read actual code — never guess patterns
• Dispatch ALL directory workers in ONE message — parallel, not sequential
• Leader NEVER writes CLAUDE.md files directly — always delegates to workers
• Use AskUserQuestion for ALL user interactions — never ask questions in plain text
• Do NOT auto-commit — let the user review generated files

Step 1: Detect Stack — MANDATORY

Run: echo "[context:1] detecting project stack"

Read references/stack-detection.md (bundled with this skill) before starting. It contains the full detection tables for languages, frameworks, test runners, asset pipelines, and databases (sections 1a–1f).

Follow sections 1a–1f from that reference, then echo the output summary as shown in section 1f.

Step 2: Read Config — MANDATORY

Run: echo "[context:2] reading blueprint config"

Read blueprint/.config.yml for:

• staging_branch — used in root CLAUDE.md branch flow
• language — content language
• stack — compare against auto-detected values; flag discrepancies

Step 3: Query Context7 for Framework Docs

Run: echo "[context:3] querying Context7 for framework documentation"

For the primary detected framework, use Context7 MCP tools:

3a. Resolve Library ID

Call mcp__context7__resolve-library-id with the framework name:

• Laravel → "laravel"
• Next.js → "nextjs"
• Django → "django"
• Rails → "ruby on rails"
• React → "react"
• Vue → "vue"
• FastAPI → "fastapi"
• Gin → "gin golang"
• etc.

3b. Query Documentation

Using the resolved library ID, call mcp__context7__query-docs for:

1. Directory structure — query: "project directory structure conventions"
2. Key patterns — query: "best practices and common patterns"
3. Testing — query: "testing conventions and patterns"

Use these results to inform:

• Which subdirectory CLAUDE.md files to generate
• What conventions to include in each file
• Framework-specific patterns and anti-patterns

3c. Context7 Fallback

If Context7 MCP tools are not available (tools not found, server not running):

1. Log: echo "[context:3] Context7 unavailable — using built-in conventions"
2. Fall back to common conventions for the detected framework (see Step 5 framework templates)
3. Continue without error — Context7 is optional enrichment, not required

Step 4: Scan Project Structure

Run: echo "[context:4] scanning project structure"

4a. Find Existing CLAUDE.md Files

find . -name "CLAUDE.md" -not -path "*/vendor/*" -not -path "*/node_modules/*" -not -path "*/.git/*" 2>/dev/null

4b. Map Project Directories

Identify directories that exist and would benefit from a CLAUDE.md:

• Directories with 5+ files that share patterns
• Directories with non-obvious rules or conventions
• Directories with complex flows or critical gotchas
• Directories with external API integrations

4c. Determine Mode

• No existing CLAUDE.md files → Full generation mode
• Existing CLAUDE.md files found → Audit mode (see Step 7)

Step 5: Generate Root CLAUDE.md

Run: echo "[context:5] generating root CLAUDE.md"

The root CLAUDE.md should be customized with detected stack info. Structure:

# {Project Name}

## Tech Stack
{Auto-detected: language, framework + version, test runner, assets, database}

## BLUEPRINT Workflow
This project uses BLUEPRINT SDLC. Run `blueprint update` to update skills.

### Pipeline
/plan → /plan-review → /code → /test → /tdd-review → /commit → /push → /ship → /finish

### Branch Flow
feature branch → {staging_branch} → main

### Commit Format
<emoji> <type>: <description>

## Testing
- Runner: {detected test runner}
- NEVER mock what you can test — use real implementations
- {framework-specific testing notes}

## Key Conventions
{Framework-specific conventions from Context7 or built-in knowledge}

## Workspace
Plans, tasks, and context live in `blueprint/` — see README there.

Keep it concise — under 60 lines. The root file is a MAP, not an encyclopedia.

Step 6: Generate Subdirectory CLAUDE.md Files

Run: echo "[context:6] generating subdirectory CLAUDE.md files"

Dispatch ALL workers in ONE message. Each worker generates CLAUDE.md files for its directory cluster. Workers use general-purpose subagent type.

Rules for Subdirectory Files

• 10-20 lines max per file — lean and focused
• Only for directories that actually exist — never create directories
• Content from Context7 results OR built-in conventions — prefer Context7 when available
• Document the NON-OBVIOUS — skip things any developer would know
• No duplication — don't repeat what's in root CLAUDE.md

Framework-Specific Templates

Generate CLAUDE.md files based on the detected framework. Only create files for directories that actually exist in the project.

Laravel

Directory	Key Content
app/Models/	Relationships, factories, casts, scopes, never raw SQL
app/Http/Controllers/	Single responsibility, use Form Requests, resource controllers
app/Http/Requests/	Validation rules, authorize method, custom messages
app/Services/	Business logic lives here, not in controllers
app/Livewire/	Component patterns, wire:model, events, lifecycle
database/migrations/	Never migrate:fresh, always add new migrations, never modify existing
database/factories/	Factory patterns, states, relationships
tests/	Pest patterns, factories over fixtures, no mocking DB
resources/views/	Blade/Livewire patterns, component library (DaisyUI if detected)
routes/	Route naming, middleware, group patterns
config/	Never hardcode — use env(), config caching

Next.js

Directory	Key Content
app/	App Router conventions, server vs client components, layouts, loading/error
components/	Component naming, props patterns, composition
lib/	Utility functions, API clients, shared logic
public/	Static assets only, no sensitive files
tests/ or __tests__/	Testing framework, component testing patterns
styles/	CSS modules or Tailwind patterns
types/	TypeScript interfaces, shared types

Django

Directory	Key Content
apps/ or app directories	App structure, models, views, serializers
templates/	Template inheritance, block patterns
static/	Static file handling, collectstatic
tests/	pytest fixtures, factory_boy, API test patterns
core/ or config/	Settings structure, URL configuration

Go

Directory	Key Content
cmd/	Entry points, CLI structure, flag parsing
internal/	Package patterns, interfaces, dependency injection
pkg/	Public packages, API stability
api/	API definitions, proto files, OpenAPI specs
tests/ or *_test.go	Table-driven tests, test helpers, testify patterns

React Native / Expo

Directory	Key Content
app/	Expo Router, screen patterns, layouts
components/	Component patterns, StyleSheet conventions
hooks/	Custom hooks, state management
services/ or lib/	API clients, storage, utilities
__tests__/	Jest + React Native Testing Library patterns

Rails

Directory	Key Content
app/models/	ActiveRecord patterns, validations, scopes
app/controllers/	Strong params, before_actions, REST conventions
app/views/	Partials, helpers, Turbo/Hotwire if detected
spec/ or test/	RSpec/Minitest patterns, FactoryBot
db/migrations/	Migration safety, never modify existing

Generic (Unknown Framework)

For unrecognized frameworks, generate CLAUDE.md only for directories with:

• Complex patterns (5+ related files)
• Test directories
• Configuration directories
• API / integration directories

Step 7: Audit Mode (Existing CLAUDE.md Files)

Run: echo "[context:7] auditing existing CLAUDE.md files"

If CLAUDE.md files already exist, run in audit mode:

7a. Compare Against Detected Stack

• Check framework version references — flag if outdated
• Check test runner references — flag if changed
• Check convention references — flag deprecated patterns
• Check directory references — flag if directories moved or removed

7b. Check for Staleness

• Wrong framework version mentioned
• Outdated patterns or deprecated APIs
• Missing conventions for newly added directories
• References to files/directories that no longer exist
• Incorrect testing patterns

7c. Report and Suggest

Use AskUserQuestion to present findings:

• List each stale/incorrect reference
• Suggest specific updates
• Let user approve or reject each change
• Do NOT overwrite without explicit confirmation

7d. Generate Missing Files

If new directories exist that should have CLAUDE.md but don't:

• Generate them following Step 6 rules
• Report them separately from audit findings

Step 8: Audit README.md — MANDATORY

Run: echo "[context:8] auditing README.md"

Cross-reference README.md against what was learned during the scan:

• Framework/language versions — match detected versions
• Listed integrations — still present in dependencies?
• Directory structure tree — matches actual structure?
• Commands/scripts — still valid?
• Missing sections for major patterns discovered

Use AskUserQuestion to offer README.md updates if staleness is found.

Step 9: Report

Run: echo "[context:9] generation complete"

Show summary:

• Stack detected: language, framework, test runner, assets, DB
• Context7: whether it was used, which docs were queried
• Created: list of new CLAUDE.md files generated
• Updated: list of existing CLAUDE.md files that were updated (audit mode)
• Unchanged: list of CLAUDE.md files that are still current
• Skipped: directories that exist but don't need CLAUDE.md
• README.md: staleness findings if any

Remind the user: "Review the generated files. Run /bp-commit when satisfied."

Flags

• --force / -f — Regenerate all CLAUDE.md files (overwrite existing)
• --dry-run / -d — Show what would be created/updated without writing
• --root / -r — Root CLAUDE.md only
• --audit / -a — Audit only, no new generation
• --no-context7 — Skip Context7 queries, use built-in conventions only

Use $ARGUMENTS as a specific directory path or flag.