AI User Guide

Precise context strategy

Optimization tips

• Specify the file path: better than a vague description
• Clarify the scope of the feature: more focused than broadly saying "the project has issues"
• Remove polite phrasing: do not set an expert role; state the task directly
• Use summaries: Use summaries rather than full text for long documents
• Monitoring usage: Regularly check token usage statistics to identify waste

Cost optimization example

Give AI an "exit"

Sometimes AI does not know the answer but will forcefully make one up. Tell it: "If you're unsure, say so clearly and wait for my confirmation instead of making something up."

Five standard steps

How to avoid rework

• •Explore before coding: Start writing code after understanding the project structure
• •Plan the implementation steps: List the steps to have AI execute according to plan
• •Stage-by-stage validation: Verify after each step; do not wait until the end
• •Submit on time: commit after completing each feature, making rollback easier

High-frequency commit strategy

Whenever you complete the development of an independent feature, or after fixing a bug and verifying it passes, automatically run git commit to commit the code.

Commit Message conventions

Use a concise Chinese commit message that clearly describes the changes in this commit.

You can roll back at any time

If the code breaks, you can always roll back to the previous stable version.

4.1 Cursor configuration

# Code Standards
- Do not use the `any` type
- `pnpm` must be used as the package manager
- All components must use TypeScript
- Follow the ESLint and Prettier configuration
- Use functional components and Hooks
- All API calls must include error handling

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "your-token"
      }
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem"],
      "env": {
        "ALLOWED_DIRECTORIES": ["/path/to/allowed"]
      }
    }
  }
}

4.2 Claude Code configuration

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "your-token"
      }
    },
    "sequential-thinking": {
      "command": "node",
      "args": ["path/to/sequential-thinking-server.js"]
    }
  }
}

4.3 Codex CLI configuration

First: provide the complete error log

Beginners often get scared when they see a screen full of red errors and end up copying only the last line. But AI is like a doctor: it needs to see the full symptoms to make an accurate diagnosis.

Second: loop repair mode

If AI doesn't fix it on the first try, don't give up. Describe the results after your attempts and keep AI trying.

Common error types and handling