The world's first product-side Spec-Driven Development framework for AI-native teams. 从需求到上线，一套可读、可开发、可测试、可运营的产研交付规格。
AI Delivery Spec / AI 产研交付规格 is a tool-agnostic delivery standard for product managers, AI product leads, engineering teams, and QA teams. It works with ChatGPT, Claude, Gemini, Codex, Cursor, Copilot, OpenClaw, and any AI tool that can read Markdown.
AI coding agents write code fast, but product delivery is still chaos: PRD missing acceptance criteria, prototypes not testable, dev handoff ambiguous, AI features without runtime governance.
AI 编程智能体写代码很快，但产品交付仍然是混乱的： PRD 缺验收标准、原型不可测试、交接模糊、AI 功能无运行时治理。
AI Delivery Spec gives your team a shared delivery protocol — not templates, but a routing-driven runtime that loads only what each artifact needs.
| Feature | ai-delivery-spec | Other PM Skills | spec-kit |
|---|---|---|---|
| L0-L3 delivery tiers (scale-adaptive) | ✅ | ❌ | ❌ |
| Replaceable domain modules | ✅ | ❌ | ❌ |
| Prototype testability rules | ✅ | ❌ | ❌ |
| 0D triage routing (TIER×AI×WORKFLOW) | ✅ | ❌ | ❌ |
| FRR 16-section delivery record | ✅ | ❌ | ❌ |
| Product-side spec (PRD + prototype) | ✅ | Partial | ❌ |
| Dev-side spec (code generation) | ❌ | ❌ | ✅ |
💡 Complementary with github/spec-kit: spec-kit handles spec→code, ai-delivery-spec handles requirement→spec+prototype.
| Persona | Start Here | Typical Outcome |
|---|---|---|
| Solo PM + AI agent | L0/L1, references/templates/prd-light-template.md, examples/ |
turn messy ideas into a readable PRD draft |
| 2-8 person ToB product team | L1/L2, PRD + prototype + acceptance gates | align PM, frontend, backend, algorithm, and QA before build |
| Enterprise delivery team | L2/L3, readiness, domain modules, AI/runtime governance | support bid, customer demo, regulated launch, and acceptance |
Use spec-kit when you already have an approved spec and need code-task decomposition. Use ai-delivery-spec when the requirement, prototype, domain logic, role path, or acceptance evidence is not yet clear. For AI-assisted delivery, use ai-delivery-spec first, then hand the stabilized spec to spec-kit or your coding agent.
- Human-readable PRDs for product, frontend, backend, algorithm, and QA teams.
- Embedded engineering contracts for AI-assisted development.
- Replaceable domain modules for CRM, traffic safety, and education IT.
- A single lifecycle bridge:
Discover -> Specify -> Plan -> Tasks -> Build/Verify -> Launch -> Learn/Retire.
You only need an AI tool that can read Markdown and the files in this repo. No vendor-specific runtime is required.
-
Write a minimum PRD / 写一个最小 PRD
Use AI Delivery Spec. Mode=Lite, Tier=L1. Write a light PRD for: [feature + target user + business goal]. Use prd-light-template and list missing decisions at the end. -
Review the PRD / 让 AI 检查它
Review this PRD with AI Delivery Spec Gate 1 and Gate 3. Check user story, role path, visible result, domain result, exceptions, and whether developers/QA can implement and test it. -
Upgrade to L2 when it will guide development / 升级到开发交付版
Upgrade this L1 PRD to L2 Standard. Add complete FRRs, state/action matrix, frontend/backend/QA handoff notes, acceptance criteria, and traceability.
# Option 1: Clone
git clone https://github.com/franklinxkk/ai-delivery-spec.git
# Option 2: Use with OpenClaw / Claude Code
# Point your agent to this repo and ask it to follow SKILL.md routing rulesUse AI Delivery Spec as the delivery standard.
First run 0D triage: [TIER] [AI] [WORKFLOW].
Load only the relevant entrypoint files.
Produce the requested artifact and end with gates, verification, gaps, and completion state.
Start with a real-world scenario:
- CRM Response Center — lead, opportunity, customer service, product feedback, contract/payment.
- Traffic Safety SaaS — regulated ToB/ToG workflows, mobile inspection, notices, hidden-danger remediation. See the complete L1 PRD sample.
- Higher-Education IT — academic affairs, student affairs, teaching systems, smart classrooms, AI learning assistants.
See examples/README.md for the full example index.
# Lightweight PRD
Use AI Delivery Spec, TIER=L0, WORKFLOW=prd.
Write a PRD for [your feature].
# Full delivery with prototype
Use AI Delivery Spec, TIER=L2, WORKFLOW=prototype.
Build an interactive HTML prototype for [your product].
# AI Native feature with runtime governance
Use AI Delivery Spec, TIER=L1, AI=native.
Spec an AI feature with runtime governance for [your scenario].
| Tier | Scope | Typical Artifacts | When to Use |
|---|---|---|---|
| L0 Lite | POC / validation | Simplified PRD + wireframe | Quick concept validation |
| L1 Standard | Standard project | Full PRD + interactive prototype + FRR | Regular feature delivery |
| L2 Full | Complex project | Full PRD + hi-fi prototype + complete FRR + acceptance matrix | Multi-stakeholder delivery |
| L3 Enterprise | Enterprise grade | Full suite + governance + multi-domain modules | Procurement / regulatory |
| Domain | File | Scope |
|---|---|---|
| Traffic Safety / 交通安全 | references/domain-traffic.md |
Regulated enterprise, vehicle, personnel, training |
| CRM / 客户经营 | references/domain-crm.md |
Lead, opportunity, customer 360, ticket, contract |
| Higher-Education Informationization / 高校教育信息化 | references/domain-education-it.md |
Academic affairs, student affairs, smart classroom |
Adding a new industry? Copy
references/domain-module-template.mdand customize.
Only 4 entrypoints, loaded on demand:
Default runtime has only four entrypoints.
SKILL.md ─────────────────────── triage, routing, gates
references/delivery-core.md ───── PRD, stories, DDD/API/data, lifecycle
references/prototype-testability.md ── prototype, mobile, interaction
references/advanced-extensions.md ── AI, SaaS, approval, reporting, global
Other reference files are detail libraries, loaded by trigger conditions only. This keeps context size small — your AI tool reads only what it needs.
其他 reference 文件是高级场景的明细库，按触发条件加载，避免大模型一次性吞下过多上下文。
| Gate | Purpose |
|---|---|
| Gate 1: Story-Path | user story → role path → visible result → domain result → test |
| Gate 2: Demo-Closed Prototype | every primary action has visible/domain outcome |
| Gate 3: PRD + Dev Contract | PRD is primary; engineering contract embedded & traceable |
| Gate 4: Acceptance Package | deliver only in-scope artifacts with verification |
Use only the stages needed by the requested artifact:
Discover → Specify → Plan → Tasks → Build/Verify → Launch → Learn/Retire
Learn/Retire is intentionally lightweight in the current runtime: use it for
minimum metric review, post-launch learning, and sunset planning. For causal
experiments, advanced A/B testing, or complex deprecation economics, pair this
repo with your analytics or experimentation framework and record the boundary
as an explicit gap. Start from
post-launch-review-template.md
when you only need a minimum review artifact.
py scripts/validate_skill_consistency.py
py scripts/validate_routing_scenarios.py
py scripts/validate_prd_quality.py path\to\prd.docx --manifest path\to\manifest.jsonWorks with: Claude Code • Claude Desktop • ChatGPT • Gemini • Codex • Cursor • Copilot • OpenClaw • Any AI tool that can read Markdown
- Pure code syntax debugging
- Copy rewriting
- Loose brainstorming with no delivery intent
Apache-2.0 — use freely in commercial projects.
See CONTRIBUTING.md. PRs welcome! 🎉