A Medical Professional's AI Programming Journey: A Pitfall-Avoidance Guide from Coze to Cursor

Introduction: A Medical Content Creator's AI Programming Journey

This is a real-world experience shared by a medical content creator. Starting from the beginning of this year with zero programming background, he independently developed multiple practical tools, including a 3D medical imaging system, a WeChat public account formatting tool, and a fracture heatmap generator. The pitfalls encountered and lessons learned along the way are invaluable for non-technical professionals looking to get started with AI programming.

Starting with Coze: The Best Choice for Zero-Background Lightweight App Development

Advantages of Web-Based Programming Platforms

The author initially used Coze to develop several practical tools: a public account formatting assistant, a literature-to-article converter, a density statistics tool, and more. These are all "lightweight apps" — with clear functionality, simple logic, and development achievable through conversation alone.

Background on Coze: Coze is an AI application development platform launched by ByteDance in late 2023, positioned as a "no-code/low-code" AI workflow building tool. Its core philosophy is to let ordinary users quickly build applications based on large language models through drag-and-drop components and natural language descriptions. The platform features a built-in plugin marketplace, knowledge base management, and workflow orchestration. Users can publish conversational AI Bots or automated workflows without understanding underlying API calls or server deployment. The technical essence of such platforms is the "encapsulation and orchestration" of LLM capabilities — lowering the barrier to entry while sacrificing flexibility and controllability.

For zero-background users, the core advantage of web-based programming platforms is: AI takes full control of the development process, and users only need to describe requirements in natural language. You don't need to understand code or file structures — as long as you can clearly express "what I want," the platform generates usable tools for you.

Fatal Flaws of Web-Based Programming Platforms

However, when the author attempted to develop a 3D medical imaging system using Coze, the problems became glaringly obvious:

1. Chaotic file management: All code is automatically lumped together with no modular management possible
2. One change affects everything: Modifying one area can trigger simultaneous changes across multiple other files
3. Extremely low controllability: AI modifies 3-7 files at once, and users can't determine which changes are necessary
4. Terrifying token consumption: When using high-end models like Zhipu 5.1, a single query can cost 2-5 RMB
5. Non-reusable features: Developed functional modules cannot be easily migrated to other projects

Understanding Token Consumption: A token is the basic unit of measurement for how large language models process text — roughly understood as "word fragments." In English, approximately 4 characters equal 1 token; in Chinese, approximately 1-2 characters equal 1 token. LLM API calls are typically billed based on total input tokens + output tokens. For high-end models like Zhipu GLM-4, the price per million tokens can reach tens of RMB. When an AI programming platform reads and modifies 3-7 files in a single operation, the code stuffed into the context window can amount to tens of thousands of tokens, causing per-call costs to skyrocket. This is the fundamental reason why complex projects on web-based platforms have uncontrollable development costs — behind every "let AI fix this for me" lies a hidden bill.

The core issue is: web-based platforms hand all authority to AI, and users lose control over the project. When project complexity exceeds a certain threshold, this "fully managed" mode collapses.

The Path Forward: Controlled Development with Cursor/Trae Local IDEs

Why You Need to Transition to a Local Development Environment

When lightweight apps could no longer meet his needs, the author transitioned to local IDE tools like Cursor. The core advantages of these tools are:

• You can specify development on a particular folder or file
• Clear project structure for easy management and maintenance
• Reusable functional modules with cross-project migration support
• Users retain higher authority, with AI serving only as the executor

Technical Background on AI-Native IDEs: Cursor is an AI-native code editor launched by Anysphere in 2023, deeply modified from VS Code with built-in capabilities to call mainstream models like GPT-4 and Claude. The most fundamental difference from web-based platforms is that Cursor positions AI as a "Copilot" rather than a "driver" — developers always maintain control over file structure, version control, and execution authority. Core features include: @file precise context injection, Composer multi-file collaborative editing, and semantic search based on the codebase. These tools represent a new paradigm of "human-AI collaborative programming": AI handles code generation and modification, while humans handle review, decision-making, and architecture design.

Current mainstream AI programming IDEs include: Cursor, Trae (domestic/international versions), Google's Firebase, Claude Code, OpenAI Codex, and others.

Recommended Choice for Chinese Users: Trae

For users in China, the author recommends Trae. It comes in domestic and international versions:

• The international version supports ChatGPT, DeepSeek^Ad, Google models, etc.
• The domestic version supports DeepSeek, Zhipu, MiniMax, Qwen, and other Chinese models

Trae is ByteDance's Cursor-like product tailored for the Chinese market, with localized optimizations for network stability and domestic model compatibility, making it more user-friendly for Chinese users. Both versions can be installed simultaneously on your computer and switched between as needed, balancing model diversity and network stability.

Three Core Pitfall-Avoidance Lessons for Beginners in AI Programming

Lesson 1: Know Your Level and Choose the Right Tool

The biggest pitfall is blindly using advanced tools without recognizing your own skill level.

The author provides a clear learning path recommendation:

1. Beginner stage: Use web-based platforms like Coze, let AI take full control, build 1-2 usable tools first to establish confidence
2. Intermediate stage: When web-based platforms can't meet your needs, transition to local IDEs like Cursor/Trae to learn file management and modular development
3. Advanced stage: Master fundamentals of databases, servers, and framework architecture for collaborative development

Don't jump straight to advanced tools like Claude Code or Codex. Not only are they several times more expensive, but the learning curve is much steeper and extremely unfriendly to programming beginners.

Lesson 2: Don't Learn for the Sake of Learning — Let Your Needs Drive Your Education

"Don't try to learn every new concept you encounter. Don't try to master everything you see others doing."

The AI field evolves extremely fast — even monthly cycles feel too long. Something you spend a month or two learning might be replaced by a new tool next month. The right approach is:

• Choose what to learn based on your project requirements
• If an existing tool can do the job, never build from scratch
• Focus your energy on logical frameworks and ways of thinking, not specific technical details

Lesson 3: Position Yourself as a "Manager," Not an "Executor"