Requirement Writer

Generate three types of requirement documents through interactive dialogue. They form a progressive chain — each builds on the previous:

Problem Framing → SRD → PRD
(Why do this?)    (What direction?)  (How exactly?)

File loading rules

Before starting, read references/questioning-guide.md for questioning strategy and completeness scoring.

Load templates on demand based on document type:

• Problem Framing → references/problem-framing-template.md
• SRD → references/srd-template.md
• PRD → references/prd-template.md

Workflow

Track progress with this checklist (copy into conversation):

Requirements Progress:
- [ ] Phase 1: Confirm document type + initial project context
- [ ] Phase 2: Layered questioning to collect information
- [ ] Phase 3: Identify project type + confirm optional sections
- [ ] Phase 4: Completeness scoring + user confirmation
- [ ] Phase 5: Generate document
- [ ] Phase 6: Multi-role review + retrospective

To revise: loop back to Phase 2, then regenerate.

Phase 1: Confirm document type

Ask what the user wants to generate. If unsure, recommend based on project stage:

• Idea stage, not yet decided whether to build → Problem Framing
• Direction decided, need cross-team alignment → SRD
• Ready for dev handoff, need execution spec → PRD

A user can start with Problem Framing and expand to SRD/PRD later.

Phase 2: Layered questioning

Read references/questioning-guide.md for the full strategy.

Core approach — three thinking layers, not a field-by-field checklist:

1. Problem essence (open-ended): What problem? Who has it? When? → Extract WHO, WHAT Problem, WHEN signals from the answer
2. Solution direction: How to solve it? Any references? What's in/out of scope? → Extract WHAT Job, Solution, Scope signals
3. Success definition: How to measure success? What benefits? → Extract Benefits, Metrics, A/B signals

After each answer, extract as much as possible and only ask about gaps. If the user gives comprehensive answers, skip ahead.

Phase 3: Project type identification

Determine project type from collected information and activate optional sections:

Signal in user's description	Project type	Activated sections
Single page/component/feature	Standard	None extra
Multiple pages, funnel, conversion path	Cross-page flow	PRD §4.0 User Flow
Different user segments, personalization	Segmented	PRD §4.0 User Segments
Redesign, overhaul, multi-phase	Holistic optimization	SRD §5.4 Phasing

Combinations are possible. Confirm with the user before proceeding:

"This looks like a cross-page flow project with user segmentation. I'll include User Flow and User Segments sections in the PRD. Correct?"

Phase 4: Completeness scoring

Read references/questioning-guide.md §Completeness Scoring for criteria.

Apply traffic-light scoring to each required field:

• 🟢 All green → proceed to generate
• 🟡 Has yellow → inform user which fields are thin, ask if they want to supplement
• 🔴 Has red → must ask follow-up questions for red fields before generating

Phase 5: Generate document

Critical: Before generating, re-read the relevant template file to ensure structural accuracy (long conversations may push earlier context out of window):

• Problem Framing → re-read references/problem-framing-template.md
• SRD → re-read references/srd-template.md
• PRD → re-read references/prd-template.md

Generate as Markdown file. In Claude.ai, save to /mnt/user-data/outputs/. In IDE environments (Claude Code, Cursor, Kiro), save to docs/ in the project root.

Document language: match the user's conversation language. For sections with insufficient information, mark as > ⚠️ TBD — 待补充. Do not fabricate content.

Phase 6: Multi-role review + retrospective

After generating, perform two checks:

Multi-role review — scan the document from three perspectives:

• 🎯 Product: Is the problem clear? Are metrics measurable? Is scope defined?
• 🎨 Design: Are scenarios specific enough for design work? Missing states (empty, error, loading)?
• 🔧 Engineering: Are backend/API needs clear? Data sources defined? Non-functional requirements covered?

Present findings as a brief checklist. Ask if user wants to supplement.

Retrospective — briefly note:

• Which sections are thinnest and why
• What questions could have been asked earlier to get better information

In IDE environments (Claude Code, Cursor, Kiro), append retrospective to docs/review-log.md. In Claude.ai, output in conversation for the user to reference.