Interesting:

> The AI workflow was as interesting to me as the retrocomputing work itself

> A coding AI can only hold so much in its context window … The solution was to make the project’s documentation the durable memory that the AI itself lacks. SwiftII was built as 18 numbered phases, 0 through 17, each a self-contained goal with a written record of what it was for and what shipped. The key milestones: […]

> Around the phases sit roughly 20 numbered design documents, each capturing one non-obvious decision. What was chosen, what the alternatives were and how it was done. … As the codebase and documents became bigger, each session had to load more context just to get oriented, so token budget became a real workflow constraint.

Yes that’s pretty much how I’ve used AI, though phase zero and one had to be hand-coded until there was enough code for the AI to iterate off. Hand debugging and then showing the AI helps to narrow the number of errors, and how to triage for the next run.