Obviously, I'm not writing all the games myself and Claude is doing a lot of the work but I thought it might be interesting to outline the process I'm using.

I'm working in VSCode and everything is held both locally and then committed and pushed to github.
First, I need a game where there is source code of whatever type out there and this is the case for many of the seminal BBC Micro games.
Next, I tell the AI that I am thinking of implementing game AAAA and to review the Game Development Guide pdf and any other docs that might be relevant (3d-Graphics manual for Elite). I've also set up a file of MMBasic key rules that it is told to review e.g. A=A!=A$=A%, variable scope, case insensitive, avoiding reserved words and the AI is told to reference this before anything else.

The next step is to ask it, without coding anything to conduct a design review of how to build the application.
For examples of these have a look at:
https://uktailwind.github.io/AI-design-files/Exile_Port_Design_Review.html
https://uktailwind.github.io/AI-design-files/Elite_Port_Plan.html
https://uktailwind.github.io/AI-design-files/Thust_Port_Plan.html

If you do nothing else look at the exile file for a real understanding of the current power of AI.

These typically have numerous questions in the first instance and this is where my knowledge of the inner workings of MMBasic come in as I can tell the AI the option to choose or alternatively propose a test environment to get to the best approach. Typically there will be numerous test programs written to tease out the best way forward and these are run on the real H/W. The design guide gets updated as each decision is finalised and you can see this in the versions posted.
Next I will get the AI to start going through the recommended phases. I will have a PC3 connected to a com port that the AI can access and it has scripts for xmodeming code onto and off the board and I have coached it in capturing screen images to analyse for correctness. My role is to identify bugs and issues and feed them in and, most importantly, stop it going down rabbit holes which it has a serious tendency to do.
Often, it will hit roadblocks and this is when my knowledge of MMbasic comes in to get it to understand why it is seeing an issue and to coach it into the correct approach. It is an iterative process between me and the AI and although it can code much faster and more accurately than me I'm still very confident in my intellectual advantage in architecting, overseeing, problem solving and phasing in the build.
Net result is that Elite took about 16 hours of my time and I don't suppose there was any period of more than 3 or 4 minutes when I wasn't interacting with the AI to work with it to develop the final product.
Basically, This is exactly the job I was doing way back developing major applications with a team of 20 programmers
Edited 2026-09-15 23:40 by matherp

---

Hi Peter!
Thanks a lot for this in-depth look at your workflow with AI. Do you get the impression that, from one project to the next, the AI ​​has a better understanding of what you want from it—meaning it doesn't just correct itself but actually learns? And could it be that one day you’ll only have to say, "Port this program to MMBasic for this or that firmware"?

  Quote  

Do you get the impression that, from one project to the next, the AI ​​has a better understanding of what you want from it—meaning it doesn't just correct itself but actually learns?

Sort of, but not consistently. When developing something big huge context is built up but that holds tokens and is largely lost at the end (and when the AI compacts). I now get it to note all key things as development notes and to some extent these are read in a new session but not all. This is part of my "job" to get the AI to reference all the various relevant knowledge before starting something new.