Amp Ralph

Autonomous AI agent workflow using Amp's handoff. No bash loop required.

Quick Start

New Project

# Copy amp-ralph into your project
cp -r amp-ralph/.amp your-project/
cp amp-ralph/AGENTS.md your-project/  # or merge with existing

# Start Amp and create a PRD
cd your-project
amp
> Load the prd skill and create a PRD for [your feature]

Existing Project

# Merge the .amp directory
cp -r amp-ralph/.amp/skills/* your-project/.amp/skills/
# Merge settings (or copy if none exist)
cp amp-ralph/.amp/settings.json your-project/.amp/

# Merge AGENTS.md guidance (append or integrate manually)
cat amp-ralph/AGENTS.md >> your-project/AGENTS.md

Workflow

1. Create a PRD

Load the prd skill and create a PRD for [your feature description]

Answer the clarifying questions. Output: tasks/prd-[feature].md

2. Start Autonomous Execution

Load the amp-ralph skill and execute tasks/prd-[feature].md

Amp will:

1. Convert PRD to prd.json
2. Create feature branch
3. Implement stories sequentially (or in parallel where safe)
4. Use auto-handoff when context fills up
5. Commit after each story passes checks
6. Continue until all stories complete

3. Monitor Progress

• prd.json — current story status
• progress.txt — cumulative learnings
• Git history — all commits
• Thread URL — linkable Amp thread for each session

Key Differences from Ralph

Feature	Ralph	Amp-Ralph
Execution	Bash loop spawns fresh amp	Single thread with auto-handoff
Story size	Must fit one context	Can span multiple contexts
Parallelism	Sequential only	Task tool for parallel stories
Planning	None	Oracle for complex stories
Memory	Files only	Thread history + files

Files

.amp/
├── settings.json              # Auto-handoff + permissions
└── skills/
    ├── prd/SKILL.md           # PRD generation
    └── amp-ralph/SKILL.md     # Autonomous execution
tasks/                         # Generated PRDs (auto-created)
prd.json                       # Current execution state (auto-created)
progress.txt                   # Cumulative learnings (auto-created)
AGENTS.md                      # Project guidance (merge with existing)

Configuration

.amp/settings.json

{
  "amp.experimental.autoHandoff": { "context": 90 },
  "amp.permissions": [
    { "tool": "Bash", "matches": { "cmd": "*git push*" }, "action": "ask" },
    { "tool": "Bash", "matches": { "cmd": "*git commit*" }, "action": "allow" },
    { "tool": "*", "action": "allow" }
  ]
}

• autoHandoff: Automatically continues in new thread at 90% context
• permissions: Allows commits, asks before push, allows all other tools

Requirements

• Amp CLI installed and authenticated
• Git repository

Acknowledgements

Prior Art

• Forked initially from snarktank/ralph
• Amp handoff feature: ampcode.com/manual#handoff
• Effective harnesses for long-running agents — Anthropic, Nov 26, 2025

Articles

• Ralph Wiggum as a "software engineer" — Geoffrey Huntley, Jul 14, 2025
• Ryan Carson's in-depth article on how he uses Ralph — Ryan Carson

Videos

• Ship working code while you sleep with the Ralph Wiggum technique — Matt Pocock, Jan 5, 2026
• "Ralph Wiggum" AI Agent will 10x Claude Code/Amp — Greg Isenberg / Ryan Carson, Jan 8, 2026