> Isn't this true of any greenfield project?

That is a good point and true to some extent. But IME with AI, both the initial speedup and the eventual slowdown are accelerated vs. a human.

I've been thinking that one reason is that while AI coding generates code far faster (on a greenfield project I estimate about 50x), it also generates tech-debt at a hyperastonishing rate.

It used to be that tech debt started to catch up with teams in a few years, but with AI coded software it's only a few months into it that tech debt is so massive that it is slowing progress down.

I also find that I can keep the tech debt in check by using the bot only as a junior engineer, where I specify precisely the architecture and the design down to object and function definitions and I only let the bot write individual functions at a time.

That is much slower, but also much more sustainable. I'd estimate my productivity gains are "only" 2x to 3x (instead of ~50x) but tech debt accumulates no faster than a purely human-coded project.

This is based on various projects only about one year into it, so time will tell how it evolves longer term.

▲

unlikelymordant 2 hours ago | parent [-]

In your experience, can you take the tech debt riddled code, and ask claude to come up with an entirely new version that fixes the tech debt/design issues you've identified? Presumably there's a set of tests that you'd keep the same, but you could leverage the power of ai in greenfield scenarios to just do a rewrite (while letting it see the old code). I dont know how well this would work, i havn't got to the heavy tech debt stage in any of my projects as I do mostly prototyping. I'd be interested in others thoughts.

	▲	fauigerzigerk 19 minutes ago \| parent \| next [-]
		Perhaps it depends on the nature of the tech-debt. A lot of the software we create has consequences beyond a paticular codebase. Published APIs cannot be changed without causing friction on the client's end, which may not be under our control. Even if the API is properly versioned, users will be unhappy if they are asked to adopt a completely changed version of the API on a regular basis. Data that was created according to a previous version of the data model continues to exist in various places and may not be easy to migrate. User interfaces cannot be radically changed too frequently without confusing the hell out of human users.
	▲	jjav an hour ago \| parent \| prev [-]
		> ask claude to come up with an entirely new version that fixes the tech debt/design issues you've identified? I haven't tried that yet, so not sure. Once upon a time I was at a company where the PRD specified that the product needs to have a toggle to enable a certain feature temporarily. Engineering implemented it literally, it worked perfectly. But it was vital to be able to disable the feature, which should've been obvious to anyone. Since the PRD didn't mention that, it was not implemented. In that case, it was done as a protest. But AI is kind of like that, although out of sheer dumbness. The story is meant to say that with AI it is imperative to be extremely prescriptive about everything, or things will go haywire. So doing a full rewrite will probably work well, only if you manage to have very tight test case coverage for absolutely everything. Which is pretty hard.