Ralph is a method. Ralph is not a specific model, assistant, prompt, nor is it an exact spec template.
Ralph is continuously looping on a given prompt with the ability to pause/edit/resume, to iteratively build the application and the spec together.
- You evolve a spec by sending it through Claude 10, 50, or
ntimes - Each iteration informs the next
- You can watch the evolution and catch regressions in real-time
- You can tune specs by observing behavior, cancelling the loop, editing the spec, and resuming
- You can delete
IMPLEMENTATION_PLAN.mdat any point and restart the loop
- Claude CLI installed and accessible in your PATH
- Terminal with 256-color support (recommended)
-
Install ralph:
brew tap agentic-metallurgy/ralph brew install ralph
-
Create a new project directory (or use an existing repo)
-
Create
specs/default.mdwith 5-10 lines describing what you'd like built -
Run ralph:
ralph
# Run with defaults (20 iterations, specs/ folder)
ralph
# Custom number of iterations
ralph --iterations 10
# Use a specific spec file
ralph --spec-file /path/to/spec.md
# Specify a different specs folder
ralph --spec-folder /path/to/specs/
# Use a custom loop prompt instead of the embedded default
ralph --loop-prompt /path/to/custom_prompt.md| Flag | Type | Default | Description |
|---|---|---|---|
--iterations |
int | 20 | Number of loop iterations to run |
--spec-file |
string | - | Override with a specific spec file |
--spec-folder |
string | specs/ |
Directory containing spec files |
--loop-prompt |
string | - | Override the embedded prompt with a custom file |
Ralph was inspired by:
- Geoffrey Huntley's article Ralph Wiggum as a Software Engineer - the original vision for iterative spec-driven development with AI
- HumanLayer's Advanced Context Engineering for Coding Agents - techniques for building effective AI-assisted development workflows