A ralph is a directory that defines an autonomous agent loop. It bundles a prompt, commands, and any files your agent needs. Ralphify is the CLI runtime that executes them.

grow-coverage/
├── RALPH.md               # the prompt (only required file)
├── check-coverage.sh      # command that runs each iteration
└── testing-conventions.md  # context for the agent

---
agent: claude -p --dangerously-skip-permissions
commands:
  - name: coverage
    run: ./check-coverage.sh
---

You are an autonomous coding agent working in a loop.
Each iteration, write tests for one untested module, then stop.

Follow the conventions in testing-conventions.md.

## Current coverage

{{ commands.coverage }}

ralph run grow-coverage     # loops until Ctrl+C

That's it. One directory. One command. The agent loops — running commands, building a fresh prompt with the latest output, and piping it to your agent. Every iteration starts with clean context and current data.

Works with any agent CLI. Swap claude -p for Codex, Aider, or your own — just change the agent field.

Why loops

A single agent run can fix a bug or write a function. But the real leverage is sustained, autonomous work — campaigns that run for hours, chipping away at a goal one commit at a time while you do something else.

Ralph loops give you:

• Incremental progress in small chunks. Each iteration does one thing, tests it, and commits. Small changes are easier to review and safer to ship.
• Fresh context every iteration. No context window bloat. The agent starts clean each loop and sees the current state of the codebase — including everything it changed last iteration.
• Continuous work toward a goal. The loop keeps running until you hit Ctrl+C or it reaches the iteration limit. Walk away, come back to a pile of commits.
• No micromanagement. Define the goal once in the prompt, tune with commands that feed live data back in. The agent figures out what to do next.
• The prompt is a tuning knob. When the agent does something dumb, add a rule. Like putting up a sign: "SLIDE DOWN, DON'T JUMP."

What people build ralphs for

Ralph	What it does
grow-coverage	Write tests for untested modules, one per iteration, until coverage hits the target
security-audit	Hunt for vulnerabilities — scan, find, fix, verify, repeat
clear-backlog	Work through a TODO list or issue tracker, one task per loop
write-docs	Generate documentation for undocumented modules, one at a time
improve-codebase	Find and fix code smells, refactor patterns, modernize APIs
migrate	Incrementally migrate files from one framework or pattern to another
research	Deep-dive into a topic — gather sources, synthesize, and build a knowledge base
bug-hunter	Run the test suite, find edge cases, write regression tests
perf-sweep	Profile, find bottlenecks, optimize, benchmark, repeat

The ralph format is intentionally simple — if you've written a skill file or a GitHub Action, you already know how it works. YAML frontmatter for config, markdown body for the prompt, {{ commands.name }} placeholders for live data.

Install

uv tool install ralphify    # recommended

Or if you don't have uv:

pipx install ralphify       # isolated install via pipx
pip install ralphify         # plain pip (use a virtualenv or --user)

Any of these gives you the ralph command.

Quickstart

Scaffold a ralph and start experimenting:

ralph scaffold my-ralph

Edit my-ralph/RALPH.md, then run it:

ralph run my-ralph           # loops until Ctrl+C
ralph run my-ralph -n 5      # run 5 iterations then stop

What ralph run does

Each iteration:

1. Runs commands — executes all commands in the ralph, captures output
2. Builds prompt — reads the RALPH.md body, replaces {{ commands.<name> }} placeholders with fresh output
3. Pipes to agent — runs the agent command with the assembled prompt on stdin
4. Repeats — goes back to step 1 with updated data

What it looks like

$ ralph run grow-coverage -n 3

▶ Running: grow-coverage
  1 command · max 3 iterations

── Iteration 1 ──
  Commands: 1 ran
✓ Iteration 1 completed (52.3s)

── Iteration 2 ──
  Commands: 1 ran
✗ Iteration 2 failed with exit code 1 (23.1s)

── Iteration 3 ──
  Commands: 1 ran
✓ Iteration 3 completed (41.7s)

──────────────────────
Done: 3 iterations — 2 succeeded, 1 failed

The ralph format

A ralph is a directory containing a RALPH.md file. Everything the ralph needs lives in that directory — scripts, reference docs, test data, whatever the agent might need.

my-ralph/
├── RALPH.md              # the prompt (required)
├── check-coverage.sh     # script (optional)
├── style-guide.md        # reference doc (optional)
└── test-data.json        # any supporting file (optional)

RALPH.md has YAML frontmatter for configuration and a markdown body that becomes the prompt:

Frontmatter field	Required	Description
agent	Yes	The agent command to run
commands	No	List of commands (name + run) whose output fills {{ commands.<name> }} placeholders
args	No	Declared argument names for {{ args.<name> }} placeholders
credit	No	Append co-author trailer instruction to prompt (default: true)

Commands run before each iteration. Their output replaces {{ commands.<name> }} placeholders in the prompt. Use them for test results, coverage reports, git history, lint output — anything that changes between iterations.

The technique

The Ralph Wiggum technique works because:

• One thing per loop. The agent picks the most important task, implements it, tests it, and commits. Then the next iteration starts fresh.
• Fresh context every time. No context window bloat. Each loop starts clean and reads the current state of the codebase.
• Progress lives in git. Code, commits, and a plan file are the only state that persists between iterations. If something goes wrong, git reset --hard and run more loops.

Read the full writeup: Ralph Wiggum as a "software engineer"

Share ralphs

Use agr to install shared ralphs from GitHub:

agr add owner/repo/grow-coverage   # install a ralph from GitHub
ralph run grow-coverage            # run it by name

Ralphs installed by agr go to .agents/ralphs/ and are automatically discovered by ralph run.

Documentation

Full documentation at ralphify.co/docs — getting started tutorial, prompt writing guide, cookbook, and troubleshooting.

Requirements

• Python 3.11+
• An agent CLI that accepts piped input (Claude Code, Codex, Aider, or your own)

License

MIT