Skip to main content

Coding with AI

Gemini CLI​

Gemini CLI is a CLI tool supported by Google to use Gemini through the terminal and utilize it for tasks like coding.

Currently, 1000 requests per day are free, making it very attractive in terms of cost, but it performs worse than other AI services, and most importantly, it often doesn't respect the existing codebase, making it difficult to use.

Current use cases for Gemini CLI:

  1. When creating a PoC or MVP from scratch: Since there's no codebase to respect, it can serve a certain role.
  2. Very simple tasks

There's also a method of connecting Gemini CLI to Claude Code as an MCP server to delegate some tasks to Gemini.1

Note

Recently, when using Gemini CLI or other AI features, it appears Google creates projects on its own.2 These projects are not manageable by users. While there hasn't been any billing yet, I'm observing closely.

Claude Code​

Claude Code is a coding tool provided by Anthropic. It performs the best among all existing AI coding tools, and the tangible effect is significant even with casual use. I'm also currently using Claude Code across the board.

Here are features I find additionally useful:

  1. ccusage
    • Tracks Claude Code usage. You can also check the current session's usage and remaining limits. While not a feature directly provided by Claude Code, it's very useful.
    • However, since Claude Code records are kept locally, costs are calculated separately when used from different devices.
  2. Subagent
    • You can save frequently used system prompts and invoke them.
    • Effective for repetitive tasks or automation. Personally, I found it more useful at work.
    • If necessary, you can share them with others or team members.
    • You can find examples in this Repository.

What I Found Effective​

  • Complex log analysis
  • Simple and clear repetitive tasks
  • Writing test code, TDD
  • Drafting documentation and explanations

What I Felt Should Be Done Manually​

  • Continuous prompt optimization
  • Complex, unique business logic with few references
  • Refactoring, structural improvements
    • Once you've done some work, you can delegate the remaining tasks to AI.
    • It's painful work, but I think doing it increases productivity.

Footnotes​

  1. https://github.com/infolabai/gemini-cli-mcp ↩

  2. https://github.com/google-gemini/gemini-cli/issues/2591 ↩