AI-Driven Software Development

Recent Developments

I haven’t been writing blog posts lately, but I’ve been steadily continuing my AI-driven software development.

At first, I used Cursor to try something similar to pair programming, but now I’ve shifted to a style where I rely on AI as much as possible using ClaudeCode.

I don’t really use tools or skills for the sake of automation; instead, I have ClaudeCode draw up a plan that allows it to operate autonomously, and I leave as much as possible to it. I adopted this approach to ensure that human instructions don’t become a bottleneck and to make the most of the AI’s knowledge.

The development process goes like this:

  1. I present what I want to do, and have it write a development plan outlining how to implement it.

  2. Based on the development plan, I have it outline the development steps.

  3. I set up a GitHub repository so development can take place there.

  4. For each development task, I have it create a branch and submit a pull request (PR) upon completion.

  5. When creating a PR, I have other AI models perform a cross-review.

  6. If any issues arise during development, I have it log them in the Issues section.

That’s how we’re proceeding.

Cross-Review Using Multiple AI Models

Currently, we’re primarily using ClaudeCode, but we’re also testing other AI models such as ChatGPT, GLM, and Gemini. As part of this effort, we’ve set things up so that coding and reviews can be handled by different AI models. It’s nothing particularly complicated—we simply taught ClaudeCode how to call other AI models and instructed it to perform cross-reviews when creating pull requests.

ChatGPT and Claude perform well enough, but they’re on the expensive side. Gemini—or rather, Antigravity—is a bit unreliable. Aside from those, I have a subscription to Ollama, and within that, I’ve taken notice of the model “GLM-5.3-Flash ” and am giving it a try. “ GLM-5.3 ” isn’t that cheap, but “GLM-5.3-Flash ” is quite affordable. Even so, I think it’s more than capable of handling coding tasks. “ GLM-5.3-Flash ” can be used on ClaudeCode, and the compatibility is good enough.

However, recently, the original Claude has lowered its costs, and I think both Sonnet and Opus have become much easier to use. So, depending on future pricing structures, I think it will depend on whether I end up using GLM more or if Claude alone will suffice.

Web App Development

Since I come from a game development background, I’m good at writing apps usingC++ , but when it comes to web apps, mobile apps, and other apps that require server-side development, I’m not as skilled. That’s why I’ve recently been aiming to reach a level where I can develop mobile apps simply by giving instructions to AI and release them to the app store.I’ve been going through the cycle of giving instructions and approving pull requests, and I’ve reached a point where I can develop apps using AI alone. However, things don’t always go as planned, and sometimes the AI produces code that’s clearly subpar. Still, since it has such a broad range of knowledge, it would be a waste not to utilize it. Rather than demanding human-like, pristine code, I’ve come to the conclusion that I’ll have to compromise—accepting that as long as the tests pass, it’s good enough—and continue developing that way.