[h14h]

In my experience, using AI coding agents need highly specific success criteria, and an easy way to verify its output against that criteria.

My biggest successes have come when I take a TDD approach. First I identify a subset of my work into a module with an API that can be easily tested, then I collaborate with the agent on writing correct test-cases, and finally I tell it to implement the module such that the test cases pass without any lint or typing errors.

It forces me to spend much more time thinking about use cases, project architecture, and test coverage than about nitty-gritty implementation details. I can imagine that in a system that evolved over time without a clear testing strategy, AI would struggle mightily to be even barely useful.

Not saying this applies to your system, but I've definitely worked on systems in the past that fit the "big ball of mud" description pretty neatly, and I have zero clue how I'd have been able to make effective use of these AI tools.