Can AI Cut Taobao Recommendation Development from a Week to Two Days?

1. Problems in the Recommendation Feed Business

Taobao's recommendation feed faces three major challenges: a large number of frequent demands that each take about a week to complete, a fragmented tech stack spanning iOS, Android, HarmonyOS, Weex, and DX that requires changes across five platforms, and low collaboration efficiency due to frequent product‑manager turnover and extensive clarification cycles.

2. WaterFlow – AI‑Powered End‑to‑End Development

WaterFlow leverages large language models to automate the whole pipeline from requirement definition to code generation and deployment. Product managers input their ideas, and the system automatically produces a clear requirement document and a set of development tasks.

3. Core Architecture

WaterFlow consists of three layers of context that guide the AI:

System Context : Fixed rules such as Git operations, user and code context retrieval, and output format.

User Context : Customizable per‑user preferences, coding style, and persona.

Code Context : Repository‑specific markdown files describing directory structures, workflows, and technology stacks.

Two main agents operate within this framework:

Central Agent : Generates requirement documents and development tasks from natural language input.

Code Agent (Codex) : Executes tasks in a cloud sandbox, creating new branches, committing code, and pushing changes without requiring local development environments.

4. Development Workflow

The workflow for each stack (frontend, backend, client, DX) follows the same pattern: create a Codex container, pull the main branch, create a new branch, let Codex develop the code, push the branch, preview the result, and submit a code review. This uniform process reduces manual setup and coordination.

# @ali/weex-waterfall
## Overview
- A React component npm package
- Provides UI and functionality for recommendation feeds
## Tech Stack
- Language: TypeScript
- Rendering: React + CSS
- Weex2.0: basic waterfall component and animation
- Ice3.0: scaffolding, debugging, building
## Commands
- Install: tnpm install
- Dev: tnpm run startWeex
- Build: tnpm run build
## Directory
src/
  index.tsx   # waterfall implementation
  cards/      # card collection
  components/ # shared components
  hooks/      # custom hooks
  utils/      # utilities
## Steps
- Pull remote branch, switch to main
- Create a new branch from main
- Modify code according to task
- Commit and push the branch

5. Results

Since its launch, WaterFlow has generated over 30 requirement documents and tasks, covering 30% of total demand. Average turnaround time dropped from one week to about 10 minutes, turning multi‑round handshakes into a single “handshake”. Development completion rates reached 90%, with more than 54 000 lines of code (Java, JavaScript, XML, etc.) automatically generated.

Key benefits include:

Significant reduction in communication overhead and faster requirement clarification.

Most tasks are fully executable by the AI, freeing engineers to focus on higher‑value work.

6. Future Plans

Short‑term goals focus on polishing workflows for all six supported tech stacks. Long‑term, the team aims to establish robust evaluation metrics for agents and introduce continuous learning (memory) so that WaterFlow can adapt to user preferences and improve over time.