name: ralph-wiggum
description: Autonomous AI coding with spec-driven development. Implements Geoffrey Huntley's iterative bash loop methodology where agents work through specs one at a time, outputting a completion signal only when acceptance criteria are 100% met.
license: MIT
metadata:
author: fstandhartinger
version: "1.0"
repository: https://github.com/fstandhartinger/ralph-wiggum
website: https://ralph-wiggum.ai

Ralph Wiggum

Autonomous AI coding with spec-driven development

What is Ralph Wiggum?

Ralph Wiggum combines Geoffrey Huntley's iterative bash loop with spec-driven development for fully autonomous AI-assisted software development.

The key insight: Fresh context each iteration. Each loop starts a new agent process with a clean context window, preventing context overflow and degradation.

When to Use This Skill

Use Ralph Wiggum when:

• You have multiple specifications/features to implement
• You want the AI to work autonomously through tasks
• You need consistent, verifiable completion of acceptance criteria
• You want to avoid context window problems in long sessions

How It Works

┌─────────────────────────────────────────────────────────────┐
│                     RALPH LOOP                              │
├─────────────────────────────────────────────────────────────┤
│  Loop 1: Pick spec A → Implement → Test → Commit → DONE    │
│  Loop 2: Pick spec B → Implement → Test → Commit → DONE    │
│  Loop 3: Pick spec C → Implement → Test → Commit → DONE    │
│  ...                                                        │
│                                                             │
│  Each iteration = Fresh context window                      │
│  Shared state = Files on disk (specs, plan, history)        │
└─────────────────────────────────────────────────────────────┘

Installation

Quick Install (via Skill Installers)

# Using Vercel's add-skill
npx add-skill fstandhartinger/ralph-wiggum

# Using OpenSkills
openskills install fstandhartinger/ralph-wiggum

Full Setup (Recommended)

For full Ralph Wiggum setup with constitution and interview:

# Tell your AI agent:
"Set up Ralph Wiggum using https://github.com/fstandhartinger/ralph-wiggum"

The agent will guide you through a lightweight, pleasant setup:

1. Quick Setup (~1 min) — Create directories, download scripts
2. Project Interview — Focus on your vision and goals (not tech details)
3. Constitution — Create a guiding document for all sessions
4. Next Steps — Clear guidance on creating specs and starting Ralph

For existing projects, the agent detects your tech stack automatically. The interview prioritizes understanding what you're building and why.

Core Concepts

1. Fresh Context Each Loop

Each iteration of the Ralph loop starts a new AI agent process. This means:
- No context window overflow
- No degradation over time
- Clean slate for each task

2. Shared State on Disk

State persists between loops via files:
- specs/ — Feature specifications with acceptance criteria
- ralph_history.txt — Log of breakthroughs, blockers, learnings
- IMPLEMENTATION_PLAN.md — Optional detailed task breakdown

3. Completion Signal

The agent outputs <promise>DONE</promise> ONLY when:
- All acceptance criteria are verified
- Tests pass
- Changes are committed and pushed

The bash loop checks for this phrase. If not found, it retries.

4. Backpressure via Tests

Tests, lints, and builds act as guardrails. The agent must fix issues before outputting the completion signal.

Usage

Creating Specifications

The key to success: Each spec needs clear, testable acceptance criteria. This is what tells Ralph when a task is truly "done."

# Feature: User Authentication

## Requirements
- OAuth login with Google
- Session management
- Logout functionality

## Acceptance Criteria
- [ ] User can log in with Google
- [ ] Session persists across page reloads
- [ ] User can log out
- [ ] Tests pass

**Output when complete:** `<promise>DONE</promise>`

Good criteria: "User can log in with Google and session persists"
Bad criteria: "Auth works correctly"

The more specific your acceptance criteria, the better Ralph performs.

Running the Loop

# Start building (Claude Code)
./scripts/ralph-loop.sh

# With max iterations
./scripts/ralph-loop.sh 20

# Using Codex CLI
./scripts/ralph-loop-codex.sh

Logging (All Output Captured)

Every loop run writes all output to log files in logs/:

• Session log: logs/ralph_*_session_YYYYMMDD_HHMMSS.log (entire run, including CLI output)
• Iteration logs: logs/ralph_*_iter_N_YYYYMMDD_HHMMSS.log (per-iteration CLI output)
• Codex last message: logs/ralph_codex_output_iter_N_*.txt

RLM Mode (Experimental)

Provide a large context file and the agent will treat it as external environment.
This is optional and experimental — it does not implement the full recursive runtime from the paper, but it does preserve all loop outputs on disk and guides the agent to query them as needed.

./scripts/ralph-loop.sh --rlm-context ./rlm/context.txt
./scripts/ralph-loop-codex.sh --rlm-context ./rlm/context.txt

RLM workspace (when enabled):
- rlm/trace/ — Prompt snapshots per iteration
- rlm/index.tsv — Index of all iterations
- logs/ — Full CLI output per iteration

Optional recursive subcalls:

./scripts/rlm-subcall.sh --query rlm/queries/q1.md

This mirrors the Recursive Language Model (RLM) idea: handle huge prompts by inspecting only the slices you need.

Two Modes

Key Principles

Let Ralph Ralph

Trust the AI to self-identify, self-correct, and self-improve. Observe patterns and adjust prompts.

YOLO Mode

For Ralph to work effectively, enable full autonomy:
- Claude Code: --dangerously-skip-permissions
- Codex: --dangerously-bypass-approvals-and-sandbox

⚠️ Use at your own risk. Only in sandboxed environments.

Links

• GitHub: https://github.com/fstandhartinger/ralph-wiggum
• Website: https://ralph-wiggum.ai
• Original methodology: Geoffrey Huntley's how-to-ralph-wiggum