# **Accelerating Taobao Feed Development with WaterFlow AI**
## Introduction
The Taobao recommendation feed previously faced **slow iteration cycles**, **diverse tech stack challenges**, and **inefficient collaboration**, often requiring a week to release a single demand.
With **WaterFlow** — an AI-driven end-to-end development practice — some requirements now go live in **just two days**, and product managers can even **self-produce and self-deliver** features.
In only a few months:
- **30+ requirements** implemented
- **54,000 lines** of automatically generated code
- Significant boosts to **development efficiency**
This article reveals **how WaterFlow works** and its impact on **collaboration models**.
---
## 1. Challenges in Taobao Feed Development
The **recommendation feed** in the Taobao app arranges products, content, and services into **real-time, swipeable cards**, tailored to user interests and scenarios.
Key parts include:
- **Homepage feed**
- **Post-purchase feed** (e.g., payment success, order details)
---
### 1.1 High Iteration Efficiency Demands
- **Many requirements, time-consuming**
Multiple innovation projects in parallel; each requires about **one week** end-to-end.
Average timeline per requirement: ~1 week
- **Multiple tech stacks to maintain**
Example: Post-purchase feed frontend involves **iOS**, **Android**, **HarmonyOS**, **Weex**, and **DX**.
A single UI change may span **five codebases**.
---
### 1.2 Low Collaboration Efficiency
Product managers rotate roles frequently. Many **different templates and recommendation reasons** make **knowledge transfer difficult**.
High alignment efforts and communication cost slow down progress.
---
### Are Industry Tools Enough?
We explored emerging 2024 AI programming tools but found them **too narrow**:
| Domain | Examples | Limitations |
|--------|----------|-------------|
| **Autonomous Agents** | Devin, OpenAI Agents | Tight Slack integration but costly DingTalk adaptation; no support for feed-specific preview/testing. |
| **AI Editors** | Cursor, GitHub Copilot | Focused on individual coding productivity; don't solve systemic feed release needs. |
---
### Rapid Prototyping Tools
Examples: **bolt.new**, **lovable**
Good for quick proofs-of-concept, but **average performance** for actual feed requirements — struggle with backend, client-side, and DX code.
---
### Summary of Tool Limitations
Current tools:
- Excel at **boosting individual productivity**
- Focus on **prototype validation**
- Lack **multi-stack collaboration capability**
We need:
- **End-to-end support** — from requirement to launch
---
## 2. WaterFlow: AI-Powered End-to-End Delivery
### **Goal**
Launch a Taobao feed requirement **with one sentence input** and preserve **all process artifacts** automatically.
Outputs include:
- Requirement documents
- Tech design docs
- Test and data reports
→ All saved as future **context** for similar tasks
---
### Target Users
- Product managers
- Developers
- Testers
- Operations staff
Working on **information feed projects**.
---
### Implementation Approach
1. Build **core pipeline**
2. Refine **AI Agents** with real-world accuracy improvements
3. Grow **context** for future tasks
---
### Results
- **30+ requirements** live in < 2 days
- Some tasks entirely by product managers
- Development time cut by >70%
---
## 3. From Natural Language to Code
### Step 1: Central Agent → Requirement Docs + Tasks
Just as TCP uses a **three-way handshake** for reliability, **product-development teams** often need multiple “N-way” clarifications before coding.
---
**Two key artifacts:**
| Type | Purpose | Generation Method |
|------|---------|-------------------|
| **Requirement Document** | Clarifies “why,” “what,” and success metrics; single source of truth. | LLM reasoning + feed domain insights |
| **Development Task** | Prompt for Code Agent — specifies **where to modify what**. | Converted from requirement doc content |
---
**Example:**
- ❌ "Add 88VIP recommendation reason to product card."
- ✅ "Add 88VIP recommendation reason **below** product card title, with **same font style as instant discount**."
---
Quote:
> *"Natural language is becoming the new programming interface, and large models will do the rest."* — Andrej Karpathy
---
### Step 2: Code Agent → Code Generation
We use **Codex** — an internal cloud-based sandbox AI coding tool.
Why cloud?
Product teams avoid **local setup** — ready-to-use for non-dev roles.
---
**Codex Components:**
- **Docker sandbox** — secure, isolated repo environment
- **LangGraph-based AI agent** — code analysis & execution tools built-in
- **Optimized LLM** — precise code generation per instructions
---
### Three Context Layers
1. **System Context** — Core WaterFlow rules (immutable)
2. **User Context** — Dev preferences & habits
3. **Code Context** — Repo-specific processes & tech docs
Example **AGENTS.md snippet** for frontend feed:
@ali/weex-waterfall
- NPM package for React component
- Tech stack: TypeScript, React, CSS, Weex 2.0
---
**Prompt vs Context:**
- **Prompt** — Instruction ("How" and "standard")
- **Context** — Background materials ("Based on what")
---
### Multi-Tech Stack Workflow Support
Supports:
- **Frontend**
- **Backend**
- **Client** (iOS/Android/Harmony)
- **DX**
Common flow: Create branch → Develop → Deploy
| Stack | Pre-Dev | Dev | Post-Dev |
|-------|--------|-----|----------|
| Frontend | Create Codex container | Pull branch, develop | Push, deploy, preview |
| Backend | Same | Same | Deploy config |
| Client | Same | Same | Compile/package |
| DX | Update template | Same | Preview new template |
---
## 4. Online Results
After 3 months:
---
### Collaboration Efficiency
- 30+ docs/tasks generated in ~10 minutes each
- Lowered bar for requirement submission
- Converted “multi-handshake” into “one-time handshake”
---
### Development Efficiency
- Codex executed tasks with **~90% completion satisfaction**
- Saves setup → build → deploy time
- Suitable for **intern-level complexity** type tasks
Examples of 100% AI-completed tasks:
- Copy improvements
- Cashback card restyle
- URL parsing config updates
- 88VIP recommendation reason addition
- Performance tracking point additions
---
**54,000+ lines of AI-generated code** across Java, JS, XML, etc.
---
### Cross-Business Efficiency Cases
| Business | Function | Role | Effect |
|----------|----------|------|--------|
| My Reviews | Add subtitle, restyle | PM, Ops | Preview & develop feeds independently |
| Admin Tools | Self-deploy feed | PM, Ops | Independent preview before production |
---
### Example: iOS Dark Mode Adaptation
#### iOS code — Background color changes
#### DX code — Font/icon adaptation
---
### Example: 88VIP Server Logic
WaterFlow handled logic generation:
---
## 5. Summary & Future Plans
**Can product managers 'develop' directly?**
→ **Partially yes** — with clearly defined tasks.
Near-term:
- Refine multi-client processes
Long-term goals:
- **Evaluation standards** — Build continuous improvement loop
- **Memory functionality** — Agents learn & personalize over time
---
## Team Intro
We are Taotian Group’s **User Terminal Technology Team**, delivering products for **hundreds of millions of Taobao users**.
We aim for:
- **Cross-platform stability**
- **Intelligent client solutions**
- **AI-powered requirement-to-release automation**
**Join us:** [Job Posting](https://talent.taotian.com/off-campus/position-detail?positionId=100001940001&shareCode=8l%2FGMUcSfnicARi9OL4aAjGlku0PpXny5Uv697sCaOGfpOa9IdOAefK4ayJCEBjs)
📺 [Video Demo](https://b23.tv/ybqP7FD)
---
