In the first half of the year I wrote a six-month Copilot report. The second half brought even bigger changes — Cursor emerged, and AI coding evolved from a "completion tool" to a "coding partner".

Copilot vs Cursor

Copilot is a code completion tool; Cursor is an AI-native editor. Their positioning is fundamentally different:

Copilot:  在你写代码时给你建议，你来决定接不接受
Cursor:   整个编辑器围绕 AI 设计，你描述意图，AI 来改代码

Cursor's Core Features

Cmd+K：自然语言编辑

Select a code block, press Cmd+K, and enter an instruction:

把这段逻辑提取成一个 custom hook，
返回 { data, loading, error, refetch }

Cursor directly modifies the code, not just gives suggestions. And it can see the project context, so the generated code will reference existing types and utility functions from the project.

Cmd+L：对话模式

Similar to ChatGPT, but with awareness of the entire project:

Q: 这个组件的性能问题在哪？
A: PostList 组件每次渲染都创建新的 filter 函数，
   建议用 useMemo 包裹。另外 db.post.findMany
   缺少分页参数，大数据量时会全量加载。

   [Generate fix]

Click "Generate fix" to directly apply the changes.

Codebase 索引

Cursor indexes the entire project codebase. Ask it "how does this project handle errors?" and it can search and summarize the global error-handling patterns.

Real-World Experience

Suitable Use Cases

Refactoring Code (Most Efficient)

选中 200 行旧组件 → Cmd+K
"重构成函数组件，用 TypeScript，添加类型定义"

10 秒后得到重构后的代码，类型定义完整。
手动调整几个细节就完成了。

Generating Repetitive Code

"根据 UserSchema 生成对应的 API CRUD handler"
"根据这个 mock 数据生成对应的测试用例"

Understanding Unfamiliar Code

选中一段没看过的库代码 → Cmd+L
"这段代码在做什么？为什么这样写？"

不适合的场景

Complex Business Logic

AI doesn't understand business context; the generated logic may look reasonable but can have subtle issues.

Architecture Decisions

Questions like "should this project use a monorepo or polyrepo" are ones AI can't answer well.

Code Requiring Precise Control

For example, security-related code and performance-critical paths — AI-generated code requires more thorough review.

Using Together with Copilot

My actual workflow:

1. Copilot 负责日常代码补全（Tab 键接受）
2. Cursor Cmd+K 做有方向的代码修改
3. Cursor Cmd+L 做代码理解、调试辅助
4. 重要代码仍然手动审查每一行

The two don't conflict; Copilot completion + Cursor deep editing is the current optimal combination.

Team Recommendations

• Don't ban AI tools; guide teams toward correct usage
• Teaching teams to "review generated code" is more important than teaching them to "use AI"
• Don't input sensitive code (keys, core algorithms) into AI tools
• Evaluate AI tool ROI: 30-50% coding speed improvement is real, but there are learning and adaptation costs
• Keep an eye on the Cursor/Copilot competitive landscape; this field is changing very rapidly

Summary

• Cursor elevates AI from "code completion" to "coding partner" level
• Cmd+K natural language editing and Cmd+L project-aware conversation are core features
• Refactoring, generating repetitive code, and understanding unfamiliar code are the best use cases
• Complex business logic and architecture decisions still require human judgment
• Combining Copilot + Cursor is the current optimal workflow