Challenges and Solutions in AI-Powered Front-End Code Generation for B2C Platforms

In the context of increasing AI penetration, the article examines core challenges and practical solutions of AI-driven front‑end code generation for B2C (D2C) scenarios.

Background: The AI innovation team at Taobao discovered that many repetitive UI tasks can be automated, promising significant efficiency gains across internal B‑end pages.

Product capabilities: The system can generate static page content, component layouts, form fields, table cell rendering, naming conventions, class names, and enforce coding standards.

Business impact: Over 5 projects have been deployed, achieving an average 15 % efficiency boost and more than 5 000 user invocations.

Key challenges and solutions:

Prompt management & evaluation cost – built a dedicated UI evaluation platform for fast scoring and versioned prompt storage.

Low image‑recognition accuracy – introduced OCR‑based translation for Chinese text, switched to Gemini/GPT‑4o models, raising accuracy to ~90 %.

Excessive prompt length causing hallucinations – split prompts by page type and use fallback generic prompts.

Too many component nodes – simplify output to essential flags (e.g., image presence, link, copyability) to improve accuracy.

Customization variance – abstracted AI capabilities into a reusable component library, reducing D2C customization cost by 80 %.

Sample JSON output from the UI parser:

[
  {"height":15,"width":54,"word":"项目管理","x":0,"y":10},
  {"height":55,"width":15,"word":"新建项目","x":93,"y":-10},
  {"height":12,"width":13,"word":"2","x":401,"y":71}
]

AI applicability & limitations: AI excels at static content generation, naming, and style enforcement, but struggles with icon recognition, complex style variations, font attributes, and business‑specific logic.

Prompt best practices: Use a README‑style format, avoid vague terms, place critical instructions at the end, and limit returned fields to necessary ones.

Overall, the case study provides a detailed technical reference for front‑end developers and AI researchers seeking to leverage large‑model capabilities in UI code automation.