RPI Pattern (Research → Plan → Implement)
The first structured workflow for AI code agents. Three sequential phases with a dedicated command per phase: (1) Research, explore the codebase to gather context, (2) Plan, write a detailed spec before touching code, (3) Implement, execute against the plan. Prevents Claude from coding before understanding the problem.
Strengths
- Prevents Claude from coding immediately, mandatory Research phase before any modification
- Spec written before implementation, misunderstanding bugs are eliminated upstream
- Implemented by all major frameworks (GSD, BMAD, Spec Kit), universal transferable pattern
- Zero install, applies to any AI runtime by structuring prompts manually
Limitations
- Overhead on micro-tasks, too much ceremony for fixing a typo or one line of CSS
- Research phase can consume many tokens on large codebases
Best for
- Any task involving modifications in an existing codebase that Claude does not yet know
- Complex features requiring understanding existing architecture before coding
- Onboarding an agent on a new project, Research maps before acting