Skip to content

Overview

Ralph Wiggum wraps the claude CLI and runs your prompt iteratively. Each round checks whether the previous one actually finished the job; the loop ends only when Claude emits an explicit completion promise.

My part

  1. 01Daily testing on real projects — reporting what was missing and what got in the way of letting the agent run on its own.
  2. 02Co-designed the loop prompts: the rules that tell the agent it may only promise completion once the work is verified.
  3. 03Ideas that shaped the roadmap, from iteration limits and resumable state to per-project configuration.
  4. 04Development itself happened on a teammate's machine — I focused on the product, the prompts and the feedback loop.
“Built at a time when AI agents still struggled to work autonomously — they'd stop halfway and declare victory. Ralph kept them honest.”

Highlights

  • Min / max iterations, custom promises, resumable state and optional session continuation.
  • Inline ratatui status bar, markdown rendering and coloured terminal output.
  • Per-project .ralph.toml with prompt prefix, suffix and system prompt.
  • Rust 2024 edition, async with Tokio.
Next project →Canopy