Stop AI from jumping ahead.
Enforce real engineering discipline – with your approval at every gate.
Quick Start • Why DevFlow? • The 5 Phases • Skills • How It Works • FAQ • 中文文档
DevFlow is a human‑gated AI development workflow that turns vague prompts into production‑ready code – without letting the AI skip design, tests, or your approval.
It integrates BDD, DDD, SDD, and TDD into a linear, phase‑gated process:
Requirements → Data Modeling → Solution Design → Development → Testing
At the end of each phase, DevFlow stops and asks for your confirmation using ask_followup_question. No surprises, no “magic” leaps.
Copy the skills/ folder into your Cursor workspace:
cp -r skills/ .cursor/skills/In Cursor, type:
/dev-flow I need a user authentication system
Answer a few questions at each phase. In about one hour, you'll have:
- ✅ Clear requirements (with Given‑When‑Then scenarios)
- ✅ A domain model and database schema
- ✅ API contracts and task breakdown
- ✅ Fully tested backend + frontend code
- ✅ A delivery report
Real example – from blank project to a running full‑stack app in ~60 minutes.
Most AI tools are too eager – they jump straight to code, ignore domain logic, skip tests, and produce “works‑on‑my‑machine” throwaways.
DevFlow is different.
| Problem | DevFlow Solution |
|---|---|
| AI writes code without understanding the domain | DDD – aggregates, invariants, state machines |
| No shared understanding of behavior | BDD – each requirement has Given‑When‑Then scenarios |
| No design, just random files | SDD – API contracts + atomic tasks before coding |
| Zero tests, untestable code | TDD – red‑green‑refactor, 100% core coverage in rigorous mode |
| AI decides everything without asking you | Mandatory human gates – after every phase, you must click “continue” |
| External dependencies block the flow | specs/humanTodo.md – AI pauses and waits for your input |
Two files keep you and the AI perfectly aligned:
| File | Purpose | Who Updates | Example |
|---|---|---|---|
specs/todo.md |
Task tracking – what's done, what's next, phase progress | AI (auto‑updated after each task/phase) | - [x] Requirements- [ ] Data Modeling |
specs/humanTodo.md |
Human dependencies – API keys, approvals, external configs | AI writes, you check off | ⬜ Provide WeChat Pay API key⬜ Confirm order cancellation rule |
No more guessing.
- Stuck? Check
humanTodo.md– the AI will even pause and ask you to complete pending items. - Want to resume after a break? Open
todo.md– you see exactly which phase or task is next.
This turns AI from a "black box" into a transparent teammate.
| Phase | Skill | Output | Human Gate |
|---|---|---|---|
| 1. Requirements | product-requirement |
REQ-*.md (BDD scenarios) |
✅ must confirm |
| 2. Data Modeling | data-modeling |
Domain model, ER diagram, sql/init.sql |
✅ must confirm |
| 3. Solution Design | solution-design |
API contracts, task list (todo.md) |
✅ must confirm |
| 4. Development | development |
src/, tests/ (TDD driven) |
✅ must confirm |
| 5. Testing | testing |
Test report, delivery report | – flow ends |
No phase can start without your explicit “yes”.
The AI physically cannot jump ahead – it usesask_followup_questionand stops.
| Skill | Responsibility |
|---|---|
dev-flow |
Orchestrates the 5‑phase flow, enforces gates, manages todo.md and humanTodo.md |
product-requirement |
Writes REQ‑*.md with BDD scenarios, version tracking |
data-modeling |
Creates domain model (aggregates, invariants), ER diagram, SQL DDL |
solution-design |
Writes technical solution, API contracts, atomic tasks |
development |
Implements code (fast mode or strict TDD), updates task status |
testing |
Runs all tests, validates against BDD scenarios, produces delivery report |
task-tracking |
Maintains specs/todo.md and specs/humanTodo.md |
All skills are atomic, self‑contained, and designed to work with Cursor’s Agent mode.
User: /dev-flow Build a simple login page
─────────────────────────────────────────────
✅ Phase 1 done → REQ-001-login.md
⏸️ DevFlow: "Continue to Data Modeling?" [Yes] [No]
User clicks Yes
─────────────────────────────────────────────
✅ Phase 2 done → Domain model (User), ER diagram
⏸️ DevFlow: "Continue to Solution Design?" [Yes] [No]
User clicks Yes
─────────────────────────────────────────────
✅ Phase 3 done → API spec (POST /login), task list
⏸️ DevFlow: "Continue to Development?" [Yes] [No]
User clicks Yes
─────────────────────────────────────────────
✅ Phase 4 done → src/ + tests/ (TDD)
⏸️ DevFlow: "Continue to Testing?" [Yes] [No]
User clicks Yes
─────────────────────────────────────────────
✅ Phase 5 done → Tests pass, delivery report generated.
🎉 Flow complete.
📋 humanTodo.md has pending items:
⬜ Provide Stripe API key
⏸️ DevFlow: "Waiting for human tasks. Complete them and reply 'done'."
specs/
├── requirements/ # REQ-*.md (BDD scenarios)
├── 002-data-model.md # Domain model, ER, SQL
├── 003-solution.md # API contracts, architecture
├── 004-test-report.md # Test results
├── delivery-report.md # Final deliverable summary
├── todo.md # Atomic task list (auto‑updated)
├── humanTodo.md # Human tasks (API keys, approvals…)
sql/
└── init.sql # Database DDL
src/ # Generated application code
tests/ # Unit & integration tests
It’s designed for Cursor (Agent mode) because it uses ask_followup_question. But the skills can be adapted to other AI coding assistants that support similar tool calls.
No. The AI is instructed to never proceed without calling ask_followup_question. This is enforced at the prompt level. You remain in control.
Those provide many individual skills. DevFlow is a single opinionated workflow that chains them together with mandatory human gates and explicit BDD→DDD→SDD→TDD methodology.
You can skip Phase 1 and manually place your own REQ-*.md files, then start from Phase 2.
We welcome new skills, improvements to existing ones, or better gate mechanisms. Please open an issue or PR.
MIT © aikeeo
Built for developers who want AI to be a teammate, not a runaway train.
```