I had been throwing around an idea for a useful tool for a few years now. I bounced ideas off of ChatGPT maybe a year or two ago, and it didn't really go anywhere. AI couldn't do what I was looking for at the time.
But time passes, and suddenly my brothers are sharing video games they had used Claude Code and Godot to fabricate. I find this fascinating, and intend to jump in myself, only some point in the future. It sounds like a bit of fun, but nothing I really have time to invest in learning it.
And so time continues to pass, when I decide to bounce those old ideas off of ChatGPT again. This was a couple of months ago now, but I am still rather pleased with what GPT shot back. It recommended I think about my intentions backwards and provided the general shape of how a different application that accomplishes the same task could work. This general shape were project design documents.
Well, I was still a novice (still am) in the AI field, so when I thought of coding my mind went to my brothers and Claude Code. So I get a 20 dollar subscription to Claude Code, and ask GPT to provide a prompt to Claude, a prompt that would be the first step of the project, the rough shape put together by GPT.
And so it went. GPT would make a prompt, I would submit it to Claude, Claude would create a summary, GPT would review and provide the next prompt. And so it continued, for a time. At first, Claude was the only agent with access to the project folder on my computer. I would manually upload edited files to GPT, along with the summary.
As time went on I discussed the project with one of my brothers, who recommended keeping a back-up. So I had GPT walk me through the process of creating a Github account, getting Claude permissions for it, and adjusting the process. I granted GPT access to Github as well, and it could finally see the files for itself. I kept Claude as my active agent, and GPT maintained the role of reviewer and gained the responsibility of being the merger. There was an addition of a minor step to have Claude update the local files to match the current branch of the Github as well, since GPT still could not access the files on my PC.
It was a decent chunk of time into this 'era' of my project that I recalled GPT *also* had Codex. So I base Codex in the same project folder as Claude, and now it can see the project as a whole. It was also at this point that Fable comes out, and I have it do a few audits of the project thus far. The process grows slightly more streamlined as well, as Claude will do the coding, and now GPT reviews, merges, updates the local files, and generates the next prompt.
It's quite the interesting endeavor, and I have a basic working version of the product to prove the process. There is still a great deal to go, especially in the area of polish. It has declared the project complete three times, and I've had to tend to correction. There often isn't too much to correct, but there is still a need. I tend to tread them as folks who might work for me. How accurate their product is depends upon how precise I am with what I want. I had some terrible organization early on because I hadn't told it how to organize, it just made a massive list.
One final note, I had to mess around with Powershell at some point as well, to update ghcli or something or other. That was around the Github creation and connection. I should also mention that the project is still kept private as it is not complete, nor will I share what the project is because I both do not wish to self-promote nor is it really relevant to the usage of AI. Well, it could be, but this project is not gardening.
I suppose I will also note that it's been around 70 million tokens for this project so far. I have no context for this usage.
This has been my experience thus far, do you have any thoughts or recommendations?
[link] [comments]