RPI in Practice: Research, Plan, Implement with Claude Code Subagents
Our first article on the RPI framework covered the why: coding agents do their best work when you split a task into Research, Plan and Implement, and keep each phase's context clean. This one is the how. It is the working setup we use day to day with Claude Code subagents, including the parts that went wrong. Nothing here is a benchmark. It is one small team running a monorepo of about a dozen…
This article explains the how of the RPI framework, focusing on the practical implementation with Claude Code subagents. It covers the setup used daily by a small team managing various products in a monorepo.
The research phase, conducted in a subagent, returns a summary or creates a file. Planning happens in the main conversation, where the user provides clear and concise instructions for the implementation phase. This phase runs in one or more subagents, each with a specific scope and clear ownership of files and tools.
Verification is crucial and should be done against something objective, such as a file on disk or a fresh page load, rather than the agent's own echo of what it did. The article also emphasizes the importance of clear directives, avoiding scope creep, and verifying results independently.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
