AI Enabled Development

Context is a scarce resource. Minimizing context consumption will enable AI (and humans for that matter) coding assistants to work on our code with much greater competency. When you design by contract and use dependency injection, you remove an entire universe of code an AI agent needs to ingest. It can iterate on the logic of your program with significantly fewer tokens consumed. This often yields faster more accurate output and enables you to use older, cheaper, or faster models with minimal downsides.

It also has the secondary benefit of reducing the complexity of your prompts. You do not need to produce large design documents to front-load your model's context window before it parses your source code. Instead your source code becomes the prompt. And your source code is infinitely more representative of your program's logic than your design document.

I have found that defining types, callbacks, and at least one exemplar implementation has enabled me to use AI as an advanced code generation tool. Not a thinking or reasoning problem-solver but a simple code generation utility which is able to translate simple, human language prompts into an accurate implementation. This keeps my problem-solving skills sharp as I'm not outsourcing my thinking, it reduces some of the menial drudgery that writing certain programs involves, and, as an unexpected side benefit, it has helped keep my wrists healthy.

All of this has not, however, improved my productivity. It has made me happier, I think, but I still deliver code at the speed of thought. It has certainly reduced the cost of experiementation but my experiments have grown in size to accommodate. Your results my vary.