solo-scaffold（单独脚手架）

/scaffold

Scaffold a complete project from PRD + stack template. Creates directory structure, configs, CLAUDE.md, git repo, and pushes to GitHub. Studies existing projects via SoloGraph for consistent patterns, uses Context7 for latest library versions.

Steps

1. Parse arguments from $ARGUMENTS — extract <project-name> and <stack-name>.

- If not provided or incomplete, use AskUserQuestion to ask for missing values. - Show available stacks from templates/stacks/*.yaml (source of truth). If MCP project_info available, also show detected stacks from active projects. List stack names with one-line descriptions from each YAML's description field. - Project name should be kebab-case.

1. Load org defaults from ~/.solo-factory/defaults.yaml:

- Read org_domain (e.g. com.mycompany), apple_dev_team, github_org, projects_dir - If file doesn't exist, ask via AskUserQuestion: - "What is your reverse-domain prefix for bundle IDs?" (e.g. com.mycompany) - "Apple Developer Team ID?" (optional, leave empty if no iOS) - Create ~/.solo-factory/defaults.yaml with answers for future runs - Replace <org_domain>, <apple_dev_team>, <github_org> placeholders in all generated files

1. Load stack + PRD + principles:

- Look for stack YAML: search for stacks/<stack>.yaml in plugin templates (via kb_search or Glob). - If stack YAML not found, use built-in knowledge of the stack (packages, structure, deploy). - Check if PRD exists: docs/prd.md or search current directory for prd.md - If not: generate a basic PRD template - Look for dev principles: search for dev-principles.md and dev-principles-my.md (personal extensions) or use built-in SOLID/DRY/KISS/TDD principles.

1. Study existing projects via SoloGraph (learn from your own codebase — critically): Before generating code, study active projects with the same stack. Don't blindly copy — existing projects may have legacy patterns or mistakes. Evaluate what's actually useful. a. Find sibling projects — use project_info() to list active projects, filter by matching stack. Example: for ios-swift, find existing projects with matching stack. b. Architecture overview — codegraph_explain(project="<sibling>") for each sibling. Gives: directory layers, key patterns (base classes, protocols, CRUD), top dependencies, hub files. c. Search for reusable patterns — project_code_search(query="<pattern>", project="<sibling>"): d. Check shared packages — codegraph_query("MATCH (p:Project)-[:DEPENDS_ON]->(pkg:Package) WHERE p.name = '<sibling>' RETURN pkg.name"). Collect package versions for reference (but verify with Context7 for latest). e. Critically evaluate what to adopt vs skip: Goal: Generated code should feel consistent with your portfolio but use the best available patterns, not just the same old ones. Limit to 2-3 sibling projects to keep research focused.

- Search for stack-specific patterns: "MVVM ViewModel", "SwiftData model", "AVFoundation recording" - Search for shared infrastructure: "Makefile", "project.yml", ".swiftlint.yml" - Search for services: "Service protocol", "actor service" - Adopt: consistent directory structure, Makefile targets, config patterns (.swiftlint.yml, project.yml) - Adopt: proven infrastructure patterns (actor services, protocol-based DIP) - Skip if outdated: old API patterns (ObservableObject → @Observable), deprecated deps - Skip if overcomplicated: unnecessary abstractions, patterns that don't fit the new project's needs - Always prefer: Context7 latest best practices over old project patterns when they conflict

1. Context7 research (latest library versions and best practices):

- For each key package from the stack: - mcp__context7__resolve-library-id — find the Context7 library ID - mcp__context7__query-docs — query "latest version, project setup, recommended file structure, best practices" - Collect: current versions, recommended directory structure, configuration patterns, setup commands - Limit to the 3-4 most important packages to keep research focused

1. Show plan + get confirmation via AskUserQuestion:

- Project path: <projects_dir>/<name> (from defaults.yaml or current directory) - Stack name and key packages with versions from Context7 - Proposed directory structure - Confirm or adjust before creating files

1. Create project directory: mkdir -p <projects_dir>/<name>
2. Create file structure based on the stack. SGR-first: always start with domain schemas/models before any logic or views. Every project gets these common files: <projects_dir>/<name>/ ├── CLAUDE.md # AI-friendly project docs (map, not manual — see Harness Engineering) ├── Makefile # Common commands (run, test, build, lint, deploy, integration) ├── README.md # Human-friendly project docs ├── docs/ │ ├── prd.md # Copy of PRD │ ├── QUALITY_SCORE.md # Domain quality grades (harness: garbage collection) │ └── ARCHITECTURE.md # Module boundaries and dependency rules ├── cli/ # CLI utility — mirrors core business logic (CLI-First Testing principle) │ └── main.ts|py # Deterministic pipeline entry point (no LLM required) ├──.claude/ │ └── skills/ # Product-specific workflow skills │ └── dev/ │ └── SKILL.md # Dev workflow skill (run, test, deploy) └──.gitignore # Stack-specific ignores CLI-First Testing: generate a cli/ directory with a stub that imports core business logic from lib/ (or equivalent). The CLI should run the main pipeline deterministically without requiring LLM, network, or UI. This enables make integration for pipeline verification. See dev-principles.md → "CLI-First Testing". .claude/skills/dev/SKILL.md — product dev workflow skill Generate a skill that teaches Claude how to work with THIS specific project. Structure: --- name: <name>-dev description: Dev workflow for <Name> — run, test, build, deploy. Use when working on <Name> features, fixing bugs, or deploying changes. Do NOT use for other projects. license: MIT metadata: author: <github_org> version: "1.0.0" allowed-tools: Read, Grep, Glob, Bash, Write, Edit --- Body should include: This makes every scaffolded project immediately Claude-friendly — new sessions get project context via the skill. MCP server (optional): If PRD indicates a data/AI/developer product, also generate MCP server stub. See templates/mcp-skills-bundle.md for the full "MCP + Skills bundle" pattern and rules for when to generate MCP. Then add stack-specific files. See references/stack-structures.md for per-stack file listings (8 stacks: nextjs, ios, kotlin, cloudflare, astro-static, astro-hybrid, python-api, python-ml).

- Stack: key packages, versions, where configs live - Commands: make dev, make test, make build, make deploy (from Makefile) - Architecture: directory structure, naming conventions, key patterns - Testing: how to run tests, where test files live, testing conventions - Common tasks: add a new page/screen, add an API endpoint, add a model

1. Generate Makefile — stack-adapted with: help, dev, test, lint, format, build, clean, deploy targets.

- Add integration target if the project has a CLI or deterministic pipeline (stub with a comment if not yet implemented) - ios-swift must also include: generate (xcodegen), archive (xcodebuild archive), open (open.xcarchive for Distribute) - The Makefile is the canonical command interface — /build and /review use make targets instead of raw commands

1. Generate CLAUDE.md for the new project:

• Project overview (problem/solution from PRD)
• Tech stack (packages + versions from Context7)
• Skills section: list available .claude/skills/ with descriptions
• Directory structure
• Common commands (reference make help)
• SGR / Domain-First section
• Architecture principles (from dev-principles)
• Harness Engineering section (from templates/principles/harness-engineering.md):

- CLAUDE.md philosophy: "this file is a map, not a manual — keep ~100 lines, point to docs/" - Harness health checklist (subset relevant to this project) - Architectural constraints: module boundaries, data validation at boundaries - Agent legibility: lint errors must include remediation instructions - Anti-patterns: don't put knowledge in Slack/Docs, don't micromanage implementation

• Do/Don't sections
• MCP Integration section (optional, if MCP tools available): Lists available MCP tools: project_code_search, kb_search, session_search, codegraph_query, project_info, web_search

1. Generate README.md — project name, description, prerequisites, setup, run/test/deploy.
2. Generate.gitignore — stack-specific patterns.
3. Copy PRD to docs/: Copy from knowledge base or generate in place.
4. Git init + first commit: cd <projects_dir>/<name> git init && git add. && git commit -m "Initial project scaffold Stack: <stack-name> Generated by /scaffold"
5. Create GitHub private repo + push: cd <projects_dir>/<name> gh repo create <name> --private --source=. --push
6. Register project + index code (optional, if MCP tools available):

- If project_code_reindex MCP tool is available, index the new project for code search: mcp__solograph__project_code_reindex(project="<name>")

1. Output summary: Project scaffolded! Path: <projects_dir>/<name> GitHub: https://github.com/<user>/<name> Stack: <stack-name> PRD: docs/prd.md CLAUDE: configured Skills:.claude/skills/dev/ (project workflow) Next steps: cd <projects_dir>/<name> <install command> # pnpm install / uv sync / etc. <run command> # pnpm dev / uv run... / etc. Then: /setup → /plan "First feature" → /build

Verification

Before reporting "project scaffolded":

1. Verify all generated files exist (ls the directory tree).
2. Run the install command (pnpm install, uv sync, etc.) — must succeed.
3. Run the dev/build command if applicable — must not error.
4. Verify git init + first commit succeeded (git log --oneline -1).
5. Verify GitHub repo creation (gh repo view or check URL).

Never say "scaffold complete" without running the install and verifying it works.

Common Issues

Stack YAML not found

Cause: Stack template missing from templates/stacks/ or not symlinked. Fix: Skill uses built-in knowledge if template not found. To fix: ensure solo-factory/templates/stacks/<stack>.yaml exists.

GitHub repo creation fails

Cause: gh CLI not authenticated or repo name already taken. Fix: Run gh auth login first. If name taken, choose a different project name.

Context7 queries fail

Cause: MCP server not running or Context7 rate limited. Fix: Skill proceeds with stack YAML versions as fallback. Context7 enhances but is not required.

org defaults missing

Cause: ~/.solo-factory/defaults.yaml not created. Fix: Run /init first for one-time setup, or skill will ask for bundle ID and team ID interactively.