Can AI Turn Product Specs into Code? Inside Alibaba’s P2C Journey

Preface

Every month before a major promotion, the team iterates both business and technology, seeking to automate code generation by treating requirements as code and finding the relationship between them.

The article focuses on the goal of automatically generating code, sharing thoughts and stage achievements during the 618 promotion.

AI Discussion

The author reflects on AI, citing the movie "AI" where a robot boy seeks humanity, and contrasts it with current AI, which lacks emotion and self‑awareness. Examples such as AlphaGo illustrate narrow AI’s strength in specific tasks but its inability to generalize without retraining.

Philosophical perspectives like John Searle’s Chinese Room experiment highlight the difference between symbol matching and true understanding.

Background

Introducing P2C (Product‑to‑Code), which aims to automatically produce code from PRD (product requirement documents).

P2C seeks to reduce communication and coding costs by standardizing requirement collection and enabling AI‑driven code generation.

Challenges

Two main technical challenges:

Obtaining a large number of natural‑language intent samples to train models, since the same requirement can be expressed in thousands of ways and implemented in many ways.

Defining a unified DSL (domain‑specific language) to describe structured requirements, similar to how 3D printing relies on standardized design files.

Derivation

Balancing comfortable requirement entry for product managers with the need for AI‑ready structured data is difficult; overly strict standardization would hinder innovation.

The proposed three‑step plan:

Structured collection of requirements to lower NLP difficulty.

Create samples linking structured requirements to standard logical expressions (providing training data for the “ProCode” part).

Use machine learning to learn from these samples and achieve the long‑term AutoCode goal.

The team plans two platforms: P2C 1.0 for precise data labeling and P2C 2.0 for open collection and innovative output.

Decomposition

Modules are broken into four parts: common components, common actions, visual code, and driving logic. The driving logic (ProCode) is expressed as expressions, actions, pipelines, and workflows.

Expressions are built from variables, operators, and pipelines; actions combine expressions; a workflow consists of actions, expressions, and triggers.

Analysis of thousands of repositories shows each module contains dozens of logical expressions, functions, and if‑statements, confirming the decomposition model.

A DSL using When + Then + Trigger can describe all business logic, preparing for P2C 2.0.

Current State

Common logical expressions are mapped to Chinese keywords, allowing operators to compose logic via a visual, keyword‑based interface.

The platform enables operators to create simple logic without writing code, using a Chinese‑oriented composition method.

Case Study

Six modules for a “hot‑sale” event were independently assembled and launched by operators.

Future Outlook

Although AI still has limitations, advances in deep learning will bring us closer to fully automated demand‑to‑code pipelines, eventually extending automation to visual design, testing, and operations.

The vision is an “intelligent chef” that transforms high‑level goals into deliverables without detailed procedural descriptions.