prd-to-engineering-spec（PRD 至工程规范）

PRD to Engineering Spec

Overview

Transform product requirements into engineering specifications so complete that developers can implement the entire system step-by-step without ambiguity, and the resulting system can be replicated or migrated without information loss.

Skill Workflow Context

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│ prd-writing-    │────►│ prd-to-         │     │ reverse-        │
│ guide           │     │ engineering-    │     │ engineering-    │
│ Write PRD       │     │ spec            │     │ spec            │
│                 │     │ [THIS SKILL]   │     │ Code→Spec       │
└─────────────────┘     └─────────────────┘     └─────────────────┘
         │
         ▼  For AI Agent products:
┌─────────────────┐
│ ai-agent-prd    │────► This skill handles Agent PRD conversion too
└─────────────────┘

Input: Complete PRD (from prd-writing-guide or ai-agent-prd) Output: Engineering specifications at replicability-grade detail

Core Principle: Validate PRD first, design second. Incomplete requirements → incomplete specs.

Quick Start

1. Validate PRD against prd-validation-checklist.md
2. For Agent systems, also validate with agent-system-spec.md
3. Generate defect report; do NOT proceed until all ❌ resolved
4. Run bash scripts/generate_spec_skeleton.sh
5. Apply the Engineering Lenses to every component
6. Fill specs using spec-templates.md
7. Validate with bash scripts/validate_spec.sh

The Engineering Lenses

Apply these seven lenses to every component during Phases 2-3. They are the engineering equivalent of prd-writing-guide's Seven Lenses.

┌──────────────────────────────────────────────────────────────────┐
│                    The Engineering Lenses                         │
├──────────────────────────────────────────────────────────────────┤
│                                                                  │
│  1. ARCHITECTURE  How does it fit? Layers, modules, comms.       │
│  2. DATA          How modeled, validated, stored, migrated?      │
│  3. CONTRACT      Interfaces? Versioning? What breaks?           │
│  4. FAILURE       How does it fail? Detect, recover, cascade?    │
│  5. SECURITY      Auth, authz, encryption, audit, secrets?       │
│  6. OPERATIONS    Deploy, configure, monitor, scale, rollback?   │
│  7. REPLICABILITY All configs, deps, assumptions documented?     │
│                                                                  │
└──────────────────────────────────────────────────────────────────┘

For every major component (service, module, API, data store, integration):

Workflow

Phase 0: PRD Validation ──────► Defect Report ──────► PRD Complete
         ↓
Phase 1: Decomposition ──► User Stories, Requirements Matrix
         ↓
Phase 2: Technical Design (Interactive) ──► Decision Log
         ↓
Phase 3: Detailed Specs ──► Engineering Lenses on every component
         ↓
Phase 4: Test Specs ──► Unit, Integration, E2E, Acceptance
         ↓
Phase 5: Task Breakdown ──► Tasks, Milestones, Risks
         ↓
Phase 6: Assembly & Replicability Review

Phase 0: PRD Validation

Goal: Ensure PRD completeness before technical work begins.

1. Review against prd-validation-checklist.md
2. Agent systems: Also validate against agent-system-spec.md §PRD Validation
3. Mark: ✅ Present | ⚠️ Unclear | ❌ Missing
4. Do NOT proceed until all ❌ resolved

Defect Report Template

# PRD Validation Report
**PRD:** [title] | **Status:** [PASS / NEEDS REVISION]

## Critical Gaps (Must Fix)
| # | Category | Issue | Impact | Resolution |
|---|----------|-------|--------|------------|

## Warnings (Should Clarify)
## Assumptions (Document explicitly with risk-if-wrong)

Phase 1: Requirements Decomposition

Goal: Break PRD into structured, implementable requirements.

User Story Format

## US-001: [Title]
**Priority:** P0/P1/P2
As a [role], I want [action], so that [benefit].
**Acceptance Criteria:** (Given/When/Then)
**Business Rules:** BR-001, BR-002
**Dependencies:** US-002

Requirements Matrix

Phase 2: Technical Design (Interactive)

Goal: Make architectural decisions with explicit user confirmation.

Decision Process

For each significant decision:

1. Present 2-3 options with pros, cons, effort, risk, and operational cost
2. Recommend with rationale
3. Wait for user confirmation
4. Record in Decision Log

Decision Log

Tech Stack

Phase 3: Detailed Specification

Goal: Specs detailed enough for implementation without questions. Apply Engineering Lenses to every component.

Templates: spec-templates.md, feature-spec-template.md

3.1 Data Model

Per entity: purpose, all fields (type, nullable, default, constraints, description), indexes with rationale, relationships with cascade rules, validation rules, lifecycle (create/update/delete behavior).

3.2 API Specification

Per endpoint: method + path, auth/authz, request (params, body, validation rules), response (structure, all status codes with conditions), business logic steps, side effects, rate limits, idempotency.

3.3 Business Logic

Per complex rule: interface (inputs/outputs/errors), algorithm in pseudocode, decision table for branching, concrete edge case examples, transaction boundaries, rollback behavior.

3.4 Security Architecture ⭐

See security-spec-guide.md. Specify:

3.5 Operations & Deployment ⭐

See operations-spec.md. Specify:

3.6 AI/Agent Components

For AI features: ai-feature-spec.md For Agent systems: agent-system-spec.md, covering:

• Agent orchestration (reasoning loop, tool dispatch, state management)
• System prompt as versioned artifact (with testing strategy)
• Skills/Tools registration and dispatch mechanism
• Memory system (working, session, long-term storage design)
• RAG pipeline (embedding → chunking → indexing → retrieval → reranking)
• Evaluation infrastructure and continuous feedback loops
• Cost model per interaction

3.7 Migration & Compatibility

If replacing existing system. See spec-templates.md §Migration:

• Data migration strategy (ETL pipeline, validation steps, rollback)
• API backward compatibility (versioning scheme, deprecation timeline)
• Feature parity matrix (old → new mapping, intentional gaps)
• Cutover plan (strategy, rollback trigger, coexistence period)

Phase 4: Test Specification

Goal: Tests that verify all requirements are met.

Traceability Matrix

Test Types

Unit: Per function—happy path, validation, edge cases, error handling. Integration: Module interactions—API contracts, DB operations, external services. E2E: Complete user journeys—critical paths with variations. Acceptance: Given/When/Then—maps to acceptance criteria. Performance: Load/stress under specified conditions—verify NFRs. Security: Auth bypass, injection, permission escalation—verify security spec.

Templates in spec-templates.md.

Phase 5: Task Breakdown

Task Template

## TASK-001: [Title]
**Implements:** FR-001, US-001 | **Estimate:** 4h
**Description:** [what to do]
**Done Criteria:** [how to verify]
**Dependencies:** TASK-000 | **Blocks:** TASK-002

Milestones

Risk Register

Phase 6: Assembly & Replicability Review ⭐

Cross-Reference Check

• Every user story → functional requirement(s)
• Every requirement → technical spec section(s)
• Every spec → test case(s)
• Every test → traces back to requirement
• All external dependencies documented with versions

Replicability Verification

The bar: Could a competent team rebuild this system from the spec alone?

• Environment: All dependencies listed with pinned versions. Build + run commands.
• Configuration: Every env var, feature flag, config param documented (type, default, range).
• Data: Schema definitions, seed data specs, migration scripts referenced.
• Infrastructure: Deploy architecture, container specs, resource requirements.
• Integrations: Every external dep has connection details, auth, failure handling.
• Business Logic: No implicit knowledge—every "obvious" rule written down.
• Security: Auth flow, permissions, encryption, secrets handling fully specified.
• Monitoring: Metrics, alerts, dashboards defined—operable from day one.
• No TODOs/TBDs: Every placeholder resolved.

Run: bash scripts/validate_spec.sh <spec_root>

Output Structure

engineering-spec/
├── 00_Overview/
│   ├── SUMMARY.md, REQUIREMENTS_MATRIX.md, DECISION_LOG.md, TECH_STACK.md
├── 01_Requirements/
│   ├── USER_STORIES.md, FUNCTIONAL_REQS.md, NON_FUNCTIONAL_REQS.md
├── 02_Technical_Design/
│   ├── ARCHITECTURE.md, DATA_MODEL.md, API_SPEC.md
│   ├── BUSINESS_LOGIC.md, AI_COMPONENTS.md (if applicable)
├── 03_Security/
│   ├── AUTH_DESIGN.md, DATA_SECURITY.md, AUDIT_SPEC.md
├── 04_Operations/
│   ├── DEPLOYMENT.md, CONFIGURATION.md, MONITORING.md, RUNBOOK.md
├── 05_Testing/
│   ├── TEST_PLAN.md, ACCEPTANCE_TESTS.md
├── 06_Implementation/
│   ├── TASK_BREAKDOWN.md, MILESTONES.md, RISKS.md, MIGRATION.md
└── SPEC_INDEX.md

Resources

• scripts/generate_spec_skeleton.sh - Create output structure
• scripts/validate_spec.sh - Validate spec completeness
• references/prd-validation-checklist.md - PRD validation checklist
• references/spec-templates.md - Specification templates
• references/feature-spec-template.md - Feature spec template
• references/ai-feature-spec.md - AI feature specification
• references/agent-system-spec.md - Agent system engineering spec ⭐
• references/security-spec-guide.md - Security architecture spec ⭐
• references/operations-spec.md - Operations & deployment spec ⭐
• references/worked-example.md - End-to-end worked example (HelpBot PRD→Spec) ⭐

Critical Reminders

1. Validate PRD first — Incomplete requirements = incomplete specs
2. Apply Engineering Lenses — Every component through all 7 lenses
3. Confirm decisions — Document rationale, not just choices
4. Trace everything — Requirement → Spec → Test → Requirement
5. Security is architecture — Not a checkbox; design it explicitly
6. Operations from day one — Deploy/monitor/debug specs are not optional
7. Replicability is the bar — If it can't be rebuilt from the spec, it's incomplete