Ralph GitHub Repository

中文 | English

Summary

This source captures the concrete snarktank/ralph repository rather than a Ralph essay. The current raw file is a refreshed README-style clipping instead of the older pinned snapshot, but it still shows the exact shell loop, prompt contracts, plugin metadata, and packaged skills that operationalize the Ralph pattern as a reusable repo asset.

Source

• Raw file: raw/ralph/snarktankralph Ralph is an autonomous AI agent loop that runs repeatedly until all PRD items are complete..md
• Translated raw file: raw/ralph/snarktankralph Ralph is an autonomous AI agent loop that runs repeatedly until all PRD items are complete..zh.md
• Original URL: https://github.com/snarktank/ralph
• Ingest date: 2026-04-13

Key Contributions

• Grounds the Ralph loop in a real repository layout: ralph.sh, prompt.md, CLAUDE.md, prd.json.example, skills/, and .claude-plugin/.
• Shows that Ralph is dual-surface by design: the same loop is packaged for both Amp and Claude Code.
• Makes the repo-distribution model explicit: Ralph is documented as copyable project files, installable skills, and a Claude Code marketplace plugin.
• Confirms that the loop's durable memory is operational, not rhetorical: branch tracking, progress archiving, prd.json state, and commit discipline are implemented in the shell script and prompt files.

Practical Implications

• Agent harnesses can be shared as runnable repository assets, not only as prose patterns.
• Repo-local skills and plugin manifests are part of the harness design because they shape how new projects adopt the loop.
• A "fresh context" loop still needs stable local memory surfaces, and this repository shows which files carry that burden in practice.

Related Pages

• Ralph Step-By-Step Guide
• Ralph Wiggum Loop Technique
• Agent Harness Design
• Agent-First Repositories
• Codex Operating Practices