Experience Sharing

Discussion
1

Announcement and Guidelines for Using Cursor

Declaration

The following content represents purely subjective personal opinions without any objective basis. Please carefully evaluate the authenticity of the content and avoid blind acceptance.

Important Note

Current artificial intelligence technology has not yet reached the level of AGI (Artificial General Intelligence). AI is essentially still an advanced pattern matching tool, therefore do not expect to achieve complex functionality with just one sentence. You need to patiently guide AI step by step to complete tasks.

Latest Updates

Cursor has released version 0.47.7 client, which adds Claude 3.7 Sonnet Max (with a nice UI design), but the 0.05 consumption per use of Claude 3.7 Sonnet Max is quite expensive. Additionally, version 0.47.6 seems to have faster differential application.

Side note: Unfortunately, Claude 3.7 Sonnet Max is only available for paid requests. I won’t be using it as I won’t pay extra! I’m proud to be frugal!

Previous Updates (0.47.5)

Cursor released version 0.47.5 client, with model mode in Agent mode. Give me some time to test this mode and I’ll update the post later.
After using it for a morning, I didn’t notice much difference. I personally still choose to stick with claude-3.7-sonnet. Perhaps I was using it too casually? Not sure, will keep observing.

Commonly Used Model Roles

  • claude-3.7-sonnet: Currently using only this model since deepseek-r1 has started charging.

GitHub Projects Using Cursor

Here are some actual project examples developed using Cursor with different models:

  1. YamlTranslater

    • Model: claude-3.7-sonnet
    • Request count: About 40 slow requests
    • Description: A YAML file conversion tool with detailed project planning and progress tracking
    • Link: BingZi-233/YamlTranslater

  2. Appwrite SDK for Java

    • Model: claude-3.5-sonnet
    • Request count: About 450 fast requests
    • Description: A Java SDK for interacting with Appwrite backend services
    • Link: BingZi-233/sdk-for-java

  3. LuckPerms Bridge Spring Boot Starter

Essential Knowledge

  1. .cursorrules File
    A document for setting project rules, typically written in Markdown syntax. However, since AI only reads content (tokens), the file format isn’t crucial - use any format as long as the content is clear.

  2. .cursor/rules File
    An upgraded version of .cursorrules allowing more detailed rule configuration. For example, you can set different AI rules for different file types (like .java and .vue files).

  3. Normal Mode vs. Agent Mode
    In Cursor’s Composer feature, you can choose between two modes:

  • Normal Mode: Suitable for quick, single code edits without deep analysis
  • Agent Mode: AI performs more intelligent analysis using reasoning and tools for complex tasks

Important Note
Many people confuse these modes’ purposes. Remember: If you need high-quality results, make sure to switch to Agent mode!

  1. Use claude-3.7-sonnet model for requirement planning (typically stored in Markdown format in the docs folder)
  2. Use claude-3.7-sonnet to list step-by-step tasks
  3. Have claude-3.7-sonnet read and execute the task list one step at a time
  4. Switch to IDEA to resolve warnings and compilation errors
  5. Direct claude-3.7-sonnet for version control commits
  6. Start new conversation and return to step 3

Writing .cursorrules

Resources

Recommended:

Reference:

.cursorrules Writing Improvement Process

  1. Select a relevant .cursorrules file from resources
  2. Locate package management file (e.g., build.gradle.kts)
  3. Choose claude-3.7-sonnet model and add context
  4. Request modifications based on package usage
  5. Iterate on improvements
  6. Complete

Note: Customizing rules requires iteration based on usage experience

About TAB Functionality

The TAB completion feature in Cursor significantly compensates for any tool functionality gaps compared to IDEs like IDEA.

About Saving Fast Requests

Fast Request Definition

Based on maximum token usage (200K tokens for claude-3.7-sonnet costs $0.02)

Speculations:

  1. Starting new conversation before token depletion costs 1 fast request
  2. Exceeding maximum tokens (200K) costs 1 additional fast request
  3. Over 25 tool calls in single response costs 1 additional fast request

Solutions:

  1. Let claude-3.7-sonnet find context (Agent mode only)
  2. Reduce new conversation frequency
  3. Use markers in .cursorrules to identify token depletion

About Automatic Testing and Problem Fixing

  • Tool must support unit testing
  • YOLO mode must be enabled

Steps

  1. Have claude-3.7-sonnet write unit tests
  2. Configure environment variables
  3. Let AI complete testing objectives efficiently

Example Prompt

Please write unit tests for [feature], run them, and fix any issues until tests pass.

Changing Terminal

Use Ctrl + Shift + P and enter:

>Terminal: Select Default Profile

Select Git Bash and restart Cursor client.

The content has been processed through a translation program. If you have any doubts, please refer to the original post.
Cursor使用经验分享 - 搞七捻三 - LINUX DO

1 Like