prp-workflow
Comprehensive PRP (Product Requirements Prompt) workflow for feature implementation. Use when user wants to implement complex features, needs a structured implementation plan, or mentions PRPs, feature implementation, or systematic development. Handles PRP generation from requirements, clarification of ambiguities, and execution for implementation. Three modes - Generation (create PRP), Clarification (resolve ambiguities), Execution (implement feature).
When & Why to Use This Skill
The PRP (Product Requirements Prompt) Workflow is a sophisticated agentic framework designed for systematic and high-quality feature implementation. It bridges the gap between initial requirements and production-ready code through a structured three-phase process: Generation, Clarification, and Execution. By leveraging parallel research agents for codebase and documentation analysis, identifying technical ambiguities through a specialized taxonomy, and enforcing rigorous validation loops during implementation, this skill ensures context-aware, robust, and architecturally consistent software development.
Use Cases
- Implementing complex new features in existing codebases while maintaining strict adherence to local patterns and architectural conventions.
- Generating AI-optimized implementation blueprints (PRPs) from vague requirements to ensure all functional and non-functional requirements are addressed.
- Automating the technical research phase by simultaneously analyzing internal code structures and external library documentation for best practices.
- Resolving underspecified project requirements through a guided, taxonomy-based clarification process to prevent costly implementation errors.
- Executing systematic code changes with built-in validation gates and automated error correction to achieve high-quality, one-pass implementation success.
| name | prp-workflow |
|---|---|
| description | | |
| Use when user | (1) wants to implement a complex feature systematically, (2) asks to "create/generate a PRP", (3) mentions "PRP workflow" or "context engineering", (4) has a feature file to implement, (5) asks to "clarify" or "review" a PRP for ambiguities, (6) asks to "execute" an existing PRP. |
| Three modes | Generation (create PRP from requirements), Clarification (resolve ambiguities before execution), Execution (implement from PRP with validation loops). |
PRP Workflow
Three-phase workflow: Generation -> Clarification (recommended) -> Execution
Mode 1: PRP Generation
Generate a comprehensive PRP from a feature request.
Process
Launch 2 parallel Opus agents for research:
Agent #1 - Codebase Research: Search for similar features/patterns in the codebase. Identify conventions (imports, naming, structure), test patterns, and architecture. Return a summary of patterns found and a list of 5-10 key files to read.
Agent #2 - External Research: Search the web for library documentation, implementation examples, best practices, and known gotchas. Return URLs, version-specific details, and recommendations.
Read key files identified by agents to build deep understanding
Clarify with user (if needed) using AskUserQuestion for patterns to follow, integration requirements, constraints
Generate PRP using the template in references/prp-template.md
Save to
plans/{XXX}-{feature-name}.md- Auto-increment from highest number in
plans/ - Filename: lowercase, hyphens, max 50 chars
- Auto-increment from highest number in
Report confidence score (1-10) and ask: clarify or execute?
Mode 2: PRP Clarification
Identify and resolve underspecified areas BEFORE implementation.
Process
Locate PRP - Ask user or search
plans/*.mdLaunch an Opus agent to analyze the PRP for ambiguities:
The agent should analyze the PRP against the coverage taxonomy in references/clarification-taxonomy.md. Mark each category (Functional Scope, Data Model, UX Flow, Non-Functional, Integration, Edge Cases, Constraints, Terminology, Completion Signals) as Clear/Partial/Missing. Return a coverage summary table and up to 5 recommended clarification questions prioritized by (Impact × Uncertainty).
Present questions to user using AskUserQuestion (max 5, up to 4 per batch):
AskUserQuestion({ questions: [ { question: "Which authentication method should the system use?", header: "Auth", multiSelect: false, options: [ { label: "OAuth 2.0 + JWT", description: "Stateless, scalable" }, { label: "Session cookies", description: "Traditional, requires storage" } ] } ] })Integrate answers into PRP:
- Add
## Clarificationssection with### Session YYYY-MM-DD - Record:
- Q: <question> → A: <answer> - Update relevant PRP sections per integration table in taxonomy reference
- Add
Report coverage summary and recommend next step
Early Exit
- No critical gaps: proceed to execution
- PRP missing: generate first (Mode 1)
Mode 3: PRP Execution
Implement a feature using an existing PRP.
Process
Load PRP - Read completely, note success criteria and validation gates
Launch an Opus agent to execute the implementation:
The agent should:
- Create a TodoWrite task list from the PRP's implementation blueprint
- Work through tasks systematically, keeping ONE task in_progress at a time
- Follow patterns from PRP exactly (don't introduce new patterns)
- Run validation commands after each major step
- Fix failures immediately (don't accumulate)
- Return: implementation summary, files modified, validation results, any deviations from PRP
Handle blockers - If agent returns with issues, help resolve
Report completion summary
Validation Philosophy
If validation fails:
- Read error carefully
- Understand root cause
- Fix underlying issue (don't patch)
- Re-run validation
- Repeat until passing
Quick Reference
File Locations
- PRPs:
plans/{XXX}-{feature-name}.md - Template: references/prp-template.md
- Taxonomy: references/clarification-taxonomy.md
Agent Summary
| Phase | Agents | Model |
|---|---|---|
| Generation | 2 parallel (codebase + external research) | Opus |
| Clarification | 1 (coverage analysis) | Opus |
| Execution | 1 (implementation) | Opus |
Key Constraints
- Clarification: max 5 questions, 4 per AskUserQuestion batch
- Execution: one task in_progress at a time
- Validate after each major step
- If CLAUDE.md exists, follow its rules
Anti-Patterns
- Skipping research to save time
- Implementing without reading full PRP
- Ignoring validation failures
- Creating new patterns instead of following existing ones