Claude Code Guide
Skills and guides for autonomous AI-assisted development with Claude Code and Ralph loops.
What This Is
Three skills you can add to any project:
prd - Create PRDs that Claude can execute autonomously
setup-claude - Initialize or audit repositories for optimal Claude Code usage
agent-browser - Headless browser automation and E2E testing
Plus guides on how I actually use Claude Code.
Quick Start
Add Skills to Your Project
npx cc-guide add-skill prd
npx cc-guide add-skill setup-claude
npx cc-guide add-skill agent-browser
npx cc-guide add-skill --all
This copies the skills to your project's .claude/skills/ folder.
Use the Skills
In Claude Code:
/prd # Create a PRD for a new feature
/setup-claude init # Set up a new repository
/setup-claude audit # Audit an existing repository
The Idea
Traditional: You code, AI helps, you iterate forever.
This way: You define the work, AI executes, you review.
Traditional: Ralph Loop:
Human → AI → Human Human → PRD → AI Loop → Human Review
↑ ↓ ↓ ↓
└──────┘ (autonomous) (complete)
The quality of your PRD determines the quality of your output.
What You Need
| Claude Code | claude.ai/claude-code | Core CLI |
| GitHub CLI | brew install gh | Git operations |
| tmux | brew install tmux | Persistent sessions |
| Ralph TUI | npm install -g ralph-tui | Task orchestration |
| Agent Browser | npm install -g agent-browser | Browser automation |
Guides
Skills
PRD Skill (/prd)
Creates self-verifying PRDs through an interview process:
- Identifies what type of task you're doing
- Extracts everything you know via brain dump
- Asks clarifying questions
- Confirms understanding
- Generates
PRD.md and prd.json
The output is a PRD that Ralph TUI can execute autonomously.
Agent Browser Skill (/agent-browser)
Headless browser automation for AI agents:
- Navigate, click, fill forms
- Take screenshots and snapshots
- E2E testing workflows
- Web scraping
Essential for any task involving UI verification.
Setup Claude Skill (/setup-claude)
Two modes:
Init - For new repositories (13 phases):
- Environment & tools check (including tmux)
- CLI discovery & authentication
- MCP configuration (context-window aware)
- Skills, subagents, rules, hooks setup
- CLAUDE.md generation
- Ralph TUI configuration
Audit - For existing repositories (8 phases):
- Environment analysis
- CLI and context window audit
- Gap analysis
- Recommendations with interactive fixes
Repository Structure
cc-guide/
├── skills/
│ ├── prd/ # PRD creation skill
│ │ ├── SKILL.md # Entry point
│ │ ├── AGENTS.md # Comprehensive reference
│ │ ├── interview/ # Interview process guides
│ │ └── categories/ # Task-type specific guidance
│ │
│ ├── setup-claude/ # Setup/audit skill
│ │ ├── SKILL.md # Entry point
│ │ ├── workflows/ # Init & audit workflows
│ │ ├── components/ # Detailed component guides
│ │ ├── reference/ # Reference material
│ │ └── templates/ # Config templates
│ │
│ └── agent-browser/ # Browser automation skill
│ └── SKILL.md # CLI reference
│
├── guides/
│ ├── my-claude-workflow.md
│ ├── ralph-loops-explained.md
│ └── best-practices.md
│
├── templates/
│ └── prompt.hbs # Ralph TUI prompt template
│
└── bin/
└── add-skill.js # npx cc-guide add-skill
Key Concepts
CLI First, MCP Last
CLIs don't consume context window. MCPs do.
Before enabling any MCP:
- Check if a CLI exists (
gh, vercel, supabase, stripe)
- Install and authenticate the CLI
- Only use MCP when there's no CLI alternative
tmux for Persistence
Ralph loops run for hours. tmux means:
- Sessions persist through SSH disconnects
- You can detach and come back later
- Multiple loops can run in parallel
Worktrees for Parallelism
Can't run two Ralph loops on the same codebase. Git worktrees fix this:
git worktree add ../project-feature-a feature-a
Each worktree can run its own tmux session with its own Ralph loop.
License
MIT - See LICENSE
The goal isn't to remove humans from development. It's to elevate humans from typists to architects.