Skip to content

Overview

Abstract

This section looks at some Claude Code setups I use.

I'll add more over time when I'm done playing with them...

Claude Code CLI God Mode

I use Claude Code CLI as my main tool.

I wanted a way to use other models via Claude Code to

  • reduce Claude Code token usage / cost (currently on Pro plan)

  • use the best tool for the job but maintain context in Claude Code

    • Gemini is good for large code bases or data given its large context window
    • ChatGPT is good for debugging

There are different ways to achieve this:

Having played with the former, I settled on the latter as it worked better

Highlights from the Reddit Community

“Gemini CLI feels like garbage… but it has that huge context window we all love. So I added instructions to CLAUDE.md to have Claude use [Gemini CLI in non‑interactive mode] when it needs to gather information about a large part of the codebase.”

“Gemini is good at explaining single, independent things … Claude is good at doing complex tasks that require a lot of memory, deep thinking, reasoning.”

https://www.reddit.com/r/ChatGPTCoding/comments/1lm3fxq/gemini_cli_is_awesome_but_only_when_you_make

This was then extended as a Claude Code command god-cli to support OpenAI Codex also https://github.com/alexsmedile/god-cli/blob/main/god-cli.md.

So a "Map repo architecture, find cause of memory leak, and propose a precise patch" prompt to Claude Code would

  • Route to Gemini → architecture map.
  • Route to Codex → debugging/diagnosis.
  • Route to Claude → minimal diff patch.

Breakthrough Method of Agile AI-Driven Development

I researched and evaluated different Agent-Based Frameworks for Software Engineering. Breakthrough Method of Agile AI-Driven Development was the one I selected as being closest to my views.

Software Engineering 1.0 Redux shows this in action with the inputs, chat history, output artifacts.

I added several features that I will merge back:

  • a security architect (Chris)
  • a vulnerability tech (Tanja) for code analysis
    • uses sub-agents to combine traditional SAST, Dependency tools with LLMs

CLAUDE.md

This CLAUDE.md addresses some of the undesirable behaviors I observed:

  1. sycophancy - I prefer my reality raw
  2. code without test - so TDD approach
  3. code without checking for existing code or being overly eager to code
  4. doing something to get the job done - but not the actual documented plan e.g.
    1. mocks or stubs
    2. instead of installing a documented needed dependency, working around it with a lesser alternative.

Takeaways

Key Takeaways

  1. By instructing Claude Code (via a god-cli Command) to route specific tasks to specific CLI tools (Google Gemini CLI, OpenAI Codex CLI), you get the best tool for the job, while saving $.
  2. Breakthrough Method of Agile AI-Driven Development (with my security additions) fits my current views of Software Engineering with GenAI.
  3. CLAUDE.md addresses some of the undesirable behaviors I observed.