RALPH Loop: The Claude Code Workflow That 10x'd a CEO's Coding Speed

Key Takeaways

• RALPH = Repetitive Autonomous Loop for PRD Handling. A structured workflow where Claude interviews, plans, then implements — rather than jumping straight to code.
• The "grill me" skill replaces plan mode: Claude interviews you relentlessly about every aspect of the feature, exploring the codebase to answer questions it can resolve itself, until both parties reach shared understanding.
• The "create plan" skill translates the interview into a detailed PRD (Product Requirements Document) — a step-by-step implementation guide with design decisions locked in before a single line of code is written.
• The "implement" skill executes the PRD autonomously. Claude follows the plan rather than improvising, which dramatically reduces correction cycles and produces consistent, reviewable output.
• Autopilot/YOLO mode chains all three skills automatically: run once, get working code. Best for well-scoped features in codebases Claude already understands. Credit: Matt Pocock on YouTube and GitHub.

Why Planning Before Coding Changes Everything

The core insight: Claude Code is most likely to produce broken or off-target code when it improvises. By forcing a structured interview phase (grill me), you surface requirements you didn't know you had. By locking them into a PRD (create plan), you give Claude a contract to execute against rather than a vague goal to interpret.

Hewitt reports that before RALPH, roughly 40% of Claude Code outputs required significant correction. After implementing the loop consistently, that fell to under 10%. The productivity gain isn't from Claude getting smarter — it's from the workflow getting more structured.

The Three Skills

• /grill-me — Interview me relentlessly about every aspect of this plan until we reach shared understanding. Walk down each branch of the design tree resolving discrepancies one by one. For each question, provide your recommended answer. Ask one question at a time. If a question can be answered by exploring the codebase, explore it instead of asking.
• /create-plan — Synthesize the interview into a full PRD with step-by-step implementation guidance. Each step should be independently verifiable. Flag any remaining ambiguities explicitly.
• /implement — Execute the PRD exactly as written. Do not deviate from the plan. If you encounter an ambiguity not covered by the PRD, stop and ask before proceeding.

Note: The above are paraphrased from Hewitt's demo. See Matt Pocock's GitHub for the original skill implementations.

Related on OpenClawDatabase

• Skills Guide — how to write and install skills in Claude Code
• Skills Database — community-contributed skills including planning workflows