project-scaffolding（项目脚手架）

Project Scaffolding

Purpose

Generate a complete CLAUDE.md and .claude/ directory structure for any project. This gives AI agents the context they need to work effectively from the first session — build commands, conventions, quality gates, and session tracking.

Stack Detection

Before scaffolding, detect the project's tech stack by scanning for manifest files:

Detection Steps

1. List files in project root: check for manifest files above
2. If package.json exists, read scripts to find build/test/lint commands
3. If pyproject.toml exists, read [tool] sections for configured tools
4. If multiple stacks detected, note all of them (monorepo pattern)

CLAUDE.md Generation

Use the template from references/claude-md-template.md and fill in detected information. Target: under 200 lines. Claude Code truncates CLAUDE.md after ~200 lines in the system prompt, so keep it lean.

1. Project Overview — read README.md or package.json description
2. Build & Run Commands — detected from manifest scripts section (keep 8-10 most-used, add "See [manifest] for full list")
3. Triggers — map file patterns to docs/skills for on-demand deep context
4. Naming Conventions — select from references/naming-conventions.md based on stack. If the table is large, move to docs/CONVENTIONS.md and reference from Triggers
5. Test Command — detected or ask user
6. Lint Command — detected or ask user

What Gets Pre-Filled

• Core Behaviors (including self-improvement and parallel session rules)
• Triggers section (generated from detected project structure)
• Test-First Development rules
• Git Commit Rules (conventional commits)
• Forbidden Files (.env, credentials, node_modules, etc.)
• Environment Variables rules

What Needs Detection or User Input

• Project name and description
• Build, test, lint, and dev commands
• Stack-specific naming conventions
• Any custom conventions the user wants

Triggers Generation

The Triggers section maps file patterns to deeper documentation, enabling on-demand context loading without bloating CLAUDE.md.

How to Generate Triggers

1. Scan project structure for distinct directories (e.g., src/, lib/, tests/, docs/)
2. Identify existing documentation files (docs/*.md, ARCHITECTURE.md, CONTRIBUTING.md)
3. Map each source directory to its most relevant doc or skill
4. Include standard patterns: Dockerfile → deployment docs, tests/ → testing docs, new files → conventions

Example Generated Triggers

| File Pattern | Load |
|---|---|
| src/api/* | Read: docs/API.md |
| src/components/* | Read: docs/COMPONENTS.md |
| tests/* | Read: docs/TESTING.md |
| Dockerfile, docker-compose* | Read: docs/DEPLOYMENT.md |
| New .ts files | Read: docs/CONVENTIONS.md |

If no docs exist yet, generate stub documentation files with TODO placeholders during scaffolding.

MEMORY.md Seeding

Seed the project's MEMORY.md with initial lessons from the detected stack. MEMORY.md is auto-loaded into every Claude Code system prompt, creating a self-improvement loop.

Location

~/.claude/projects/<project-path-with-dashes>/memory/MEMORY.md

Where <project-path-with-dashes> is the absolute project path with / replaced by - (e.g., -Users-dev-myproject).

What to Seed

1. Stack details — language version, build tool, package manager
2. Architecture patterns — module structure discovered during detection
3. Common pitfalls for the detected stack:

- TypeScript: ESM import extensions, tsconfig strictness, type-only imports - Python: virtual env activation, __init__.py requirements, type checking with mypy - Rust: borrow checker common mistakes, feature flags, workspace patterns - Go: module paths, exported vs unexported, error wrapping

1. Testing patterns — framework, mock strategy, test file locations
2. Owner preferences — any conventions the user specified during /init

Size Budget

Keep under 55 lines initially — leave room for growth as the agent learns during sessions. The 200-line system prompt limit applies to MEMORY.md too.

Directory Structure Creation

Create the following structure:

.claude/
├── commands/
│   └── parallel.md            ← multi-session coordination guide
├── reports/
│   ├── _registry.md          ← from session-management references
│   ├── _tech-debt.md         ← from session-management references
│   ├── architecture/
│   ├── bugs/
│   ├── code-review/
│   ├── config/
│   ├── dependencies/
│   ├── docs/
│   ├── features/
│   ├── integration/
│   ├── performance/
│   ├── refactor/
│   ├── security/
│   ├── testing/
│   ├── ci/
│   ├── tech-debt/
│   └── release/
└── scripts/
    └── post-edit-check.sh     ← from quality-gates hook

Steps

1. Create all 15 category directories under .claude/reports/
2. Copy registry-template.md content to .claude/reports/_registry.md (empty, ready to use)
3. Copy tech-debt-template.md content to .claude/reports/_tech-debt.md (empty, ready to use)
4. Copy post-edit-check.sh to .claude/scripts/ using cp via Bash (NOT the Write tool — Write can corrupt line endings and break the script): mkdir -p.claude/scripts && cp "$CLAUDE_PLUGIN_DIR/scripts/post-edit-check.sh".claude/scripts/post-edit-check.sh && chmod +x.claude/scripts/post-edit-check.sh If $CLAUDE_PLUGIN_DIR is unavailable, locate the plugin via find ~/.claude -name "post-edit-check.sh" | head -1 and cp from there.
5. Create .claude/commands/parallel.md with project-specific file ownership boundaries
6. Seed MEMORY.md at the auto-memory path (see MEMORY.md Seeding section above)

Hook Setup

Based on detected stack, wire up the PostToolUse hook:

1. Create or update .claude/settings.json with exactly this content (no extra arguments, no modifications):

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "bash .claude/scripts/post-edit-check.sh"
          }
        ]
      }
    ]
  }
}

Important: The command must be exactly bash.claude/scripts/post-edit-check.sh — do not add arguments like $CLAUDE_TOOL_INPUT_FILE_PATH. The script reads $CLAUDE_TOOL_INPUT from the environment automatically.

1. Do NOT modify the script — it already handles all stacks (TypeScript, Python, Rust, Go) via file extension detection. Copy it verbatim.

Interactive Flow

When running the scaffolding, interact with the user for decisions:

Step 1: Detect and Confirm

Detected stack: TypeScript (package.json + tsconfig.json)
Build: npm run build
Test: npm test
Lint: npm run lint

Does this look correct? Any changes?

Step 2: Custom Conventions

Any project-specific conventions I should add to CLAUDE.md?
For example:
- Specific directory structure rules
- Required code patterns
- Forbidden patterns or anti-patterns

Step 3: Generate and Review

1. Generate CLAUDE.md with all detected + user-provided info
2. Create .claude/ directory structure
3. Set up hooks
4. Show summary of what was created

Step 4: Confirm

Created:
- CLAUDE.md (project configuration with Triggers, under 200 lines)
- .claude/reports/ (15 category dirs + registry + tech debt tracker)
- .claude/commands/parallel.md (multi-session coordination)
- .claude/scripts/post-edit-check.sh (auto-validation hook)
- .claude/settings.json (PostToolUse hook wired up)
- MEMORY.md (seeded at ~/.claude/projects/.../memory/MEMORY.md)

Ready to start developing with session tracking, auto-validation,
and self-improvement across sessions.

Cross-References

• references/claude-md-template.md — Full CLAUDE.md template with placeholder sections
• references/naming-conventions.md — Language-specific naming tables (TS, Py, Rust, Go)
• references/git-workflow.md — Conventional commits, branching, and PR patterns