メニュー
Autonomous AI coding loop for unattended development - lets Claude work continuously on tasks while you sleep, with mandatory sandbox enforcement for safety
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Structured planning and project breakdown using proven methodologies for goals, projects, and strategic initiatives. Use when users need to create plans, break down complex projects, set milestones, estimate timelines, identify dependencies, or develop action plans. Triggers include 'help me plan,' 'create a roadmap for,' 'break down this project,' 'what are the steps to,' or 'how should I approach.'
Expert guidance for writing professional Python code following industry best practices including PEP 8 compliance, testing, type hints, error handling, and modern tooling. Use this skill when writing new Python code, refactoring existing code, setting up Python projects, implementing tests, or ensuring code quality and maintainability. Emphasizes: PEP 8, modularity, DRY principle, TDD, virtual environments (uv), and modern tooling (Ruff, Black, Mypy).
Conduct systematic scientific literature reviews by integrating multi-source searches, iterative analysis, and synthesis into well-structured reports with proper citations. Use this skill when you need to: synthesize knowledge across multiple research domains, identify research gaps, track the evolution of methodologies, evaluate evidence quality, or write comprehensive literature reviews for academic papers, grant proposals, or research projects. Key capabilities: multi-database search strategies, iterative query refinement, content analysis and summarization, thematic synthesis, citation management, quality assessment, bias detection, and structured report generation with multiple output formats.
Expert evaluation of Chemical Engineering PhD qualifying exams - review written reports, presentations, and prepare comprehensive questioning sessions to assess student readiness for doctoral research
Comprehensive scientific writing guidance for research papers, grants, and technical documentation. Covers paper structure (IMRAD), methods writing, results presentation, figure/table design, citation formatting, abstract writing, and revision responses. Use when users are writing scientific papers, formatting manuscripts, responding to reviewers, writing grant proposals, or need help with any aspect of scientific communication. Triggers include 'write a paper,' 'scientific writing,' 'format my manuscript,' 'methods section,' 'respond to reviewers,' or any research writing task.
Structured brainstorming and ideation facilitation using proven creativity techniques. Use when users want to generate ideas, explore solutions, break through creative blocks, or need facilitated ideation sessions. Triggers include requests like 'help me brainstorm,' 'generate ideas for,' 'creative solutions to,' or 'think of alternatives.'
| name | ralph-wiggum |
| description | Autonomous AI coding loop for unattended development - lets Claude work continuously on tasks while you sleep, with mandatory sandbox enforcement for safety |
| allowed-tools | ["Bash","Read","Write","Edit","Glob","Grep","Task","TodoWrite"] |
"Me fail English? That's unpossible!" - Ralph Wiggum
Ralph Wiggum Mode is an autonomous AI coding loop that enables Claude Code to work continuously on tasks without human intervention. Named after the persistently optimistic Simpsons character, this technique lets Claude work "overnight shifts" while you sleep.
Progress persists in files and git, not in the LLM's context window.
When context fills up, the loop restarts Claude with fresh context. The new instance picks up from the filesystem state (your code, git history, and task tracking files).
Use Ralph Wiggum mode for:
CRITICAL: Ralph requires a sandboxed environment.
Ralph uses --dangerously-skip-permissions which can:
Ralph will REFUSE to start unless it detects:
# Run the init script to scaffold Ralph files
./ralph-init.sh
This creates:
RALPH_PROMPT.md - The prompt fed to Claude each iterationIMPLEMENTATION_PLAN.md - Task tracking fileralph.sh - The main loop scriptralph-sandbox.sh - Docker wrapperDockerfile.ralph - Sandbox container definition# Create specs directory for requirements
mkdir -p specs
# Write your requirements
cat > specs/my-feature.md << 'EOF'
## Feature: User Dashboard
### Requirements
- Show user's recent activity
- Display usage statistics
- Allow date range filtering
EOF
# Update the implementation plan
cat > IMPLEMENTATION_PLAN.md << 'EOF'
## Tasks
- [ ] Create Dashboard component
- [ ] Add activity feed API endpoint
- [ ] Implement date range picker
- [ ] Write tests for dashboard
EOF
# Recommended: Use the Docker wrapper
./ralph-sandbox.sh
# With iteration limit
./ralph-sandbox.sh . 20
# Check progress from another terminal
tail -f ralph.log
git log --oneline -10
# See what Ralph did
git log --oneline
# Check task completion
cat IMPLEMENTATION_PLAN.md
# Verify tests pass
pytest
┌─────────────────────────────────────────────────────────────────┐
│ EXTERNAL BASH LOOP │
│ │
│ while :; do │
│ ┌─────────────────────────────────────────────────┐ │
│ │ CLAUDE SESSION │ │
│ │ │ │
│ │ 1. Read RALPH_PROMPT.md │ │
│ │ 2. Check IMPLEMENTATION_PLAN.md │ │
│ │ 3. Pick a task, implement it │ │
│ │ 4. Run tests, commit changes │ │
│ │ 5. Context fills → Session ends │ │
│ └─────────────────────────────────────────────────┘ │
│ ↓ │
│ Git commit saves progress │
│ Loop restarts with fresh context │
│ done │
└─────────────────────────────────────────────────────────────────┘
The prompt loaded each iteration. Customize this for your use case:
# Ralph Mode Instructions
You are operating in autonomous Ralph mode.
## Each Iteration
1. Read IMPLEMENTATION_PLAN.md and git log
2. Pick ONE pending task, mark it in-progress
3. Implement fully with tests
4. Mark complete, commit changes
## Rules
- Never skip tests
- Commit after each meaningful change
- If blocked, document why and continue
Persistent task tracking (survives context resets):
## Tasks
- [x] Task 1 (completed in iteration 1)
- [x] Task 2 (completed in iteration 2)
- [ ] Task 3 (in progress)
- [ ] Task 4 (pending)
The main loop with sandbox enforcement. Key features:
Docker wrapper that:
--network none)| Variable | Default | Description |
|---|---|---|
RALPH_MAX_ITERATIONS | 0 (unlimited) | Stop after N iterations |
RALPH_MODEL | sonnet | Claude model to use |
RALPH_COST_LIMIT | 50.00 | Max spend in USD (with cost monitor hook) |
ANTHROPIC_API_KEY | optional | API key (not needed with Claude Code subscription) |
Only for custom sandbox environments you've verified:
RALPH_I_KNOW_WHAT_IM_DOING=sandboxed ./ralph.sh
Break large tasks into small, atomic pieces:
## Bad
- [ ] Implement user authentication
## Good
- [ ] Create User model with email, password_hash
- [ ] Add POST /api/register endpoint
- [ ] Add POST /api/login endpoint
- [ ] Add password reset flow
- [ ] Add JWT middleware
- [ ] Write auth tests
# Watch iterations in real-time
tail -f ralph.log
# Check commits
watch -n 5 'git log --oneline -10'
# Monitor resource usage
docker stats
Install these hooks for additional safety:
# Block dangerous operations
skillz hooks install ralph-safety-check
# Track API costs
skillz hooks install ralph-cost-monitor
Check sandbox detection output. You need at least 3 checks passing:
ralph.log for errorsIMPLEMENTATION_PLAN.md to unblockRALPH_PROMPT.mdSet MAX_ITERATIONS or use the cost monitor hook:
./ralph-sandbox.sh . 20 # Stop after 20 iterations