New Skill: structured-rpi - Enhanced Research-Plan-Implement workflow

[spences10]

Summary

Enhanced RPI workflow with structured outputs between phases. Based on 12-factor agents concepts from AI That Works show.

Problem

• Monolithic prompts have 100+ instructions, models skip steps
• Plans are 1000+ lines, nobody reviews them properly
• Feedback late in workflow = low leverage
• Models jump straight to plan writing, skip design discussion

Proposed Skill

Name: structured-rpi
Plugin: claude-workflow
Triggers: "plan this feature", "design discussion", "structure outline", "phased planning"

Core Content

5-6 Step Breakdown (vs 3)

1. Generate Questions     → what's unclear?
2. Targeted Research      → guided codebase exploration
3. Design Discussion      → structured JSON: openQuestions, resolvedQuestions
4. Structure Outline      → phases with dependencies (THE review artifact)
5. Plan Writing           → detailed steps
6. Implement              → execute

Structured Outputs

// Phase 3: Design Discussion
{
  summary: string[],
  currentState: string[],
  desiredEndState: string[],
  openQuestions: [{ title, question, options, recommendation }],
  resolvedQuestions: string[]
}

// Phase 4: Structure Outline
{
  phases: [{ name, description, dependencies }]
}

Exit Conditions

if (designOutput.openQuestions.length === 0) {
  advanceToStructurePhase()
}

Key Insight

Review structure outline, not full plan. One screen, shows dependencies, easy to catch wrong approach.

References

• rework/03-rpi-enhanced.md - detailed breakdown
• AI That Works show episode on 12-factor agents