Nobody knows how to build with AI yet

The "time dialation" is real. I mostly manage these days, yet my fun projects progress faster than they ever have, because I can prompt in the 2 minutes between meetings, and come back to significant progress.

Yes, it's not faster to develop with AI if you watch it work. It's faster to develop with AI if you parallelize. Typing was never the bottleneck, but is is a now-parallelizeable part of the pipeline.

It can still be faster to develop with AI watching it work. It can legitimately introduce an entire simple fullstack change across multiple projects in my monorepo including graphql queries/mutations, typeorm repository, a service layer, and a reactnative frontend using apollo client, etc. It can do that in about 10 minutes in my local. I can't. If I turned it into a speed run event and practiced, maybe I could get it done in 10 minutes but honestly, it's a machine and I'm John Henry. Since it's using my IDE, it's using my meticulously setup and maintained local and I'm able to quickly stop it and fix any mistake it makes. Level 2 driving assist.

I have enjoyed the github copilot agent style development where someone elses computer is running everything, and I can make a request and just come back half an hour later and check on it. But this level 5 driver gets the wrong destination basically every time, and then it's another 10, 20 or even 30 minutes for it to make a minor adjustment. It doesnt understand my `yarn` scripts, it runs my tests wrong, it can't do codegen, it doesn't format or lint files, etc. I asked copilot yesterday to lint and format a PR and it took 25 minutes of agentic work lol.

> Yes, it's not faster to develop with AI if you watch it work.

It’s actually a lot faster. You read the diffs as soon as they start coming in, and immediately course correct or re-prompt when you see bad mistakes.

I don't have this experience. Watching and course correcting like this makes me realize I could have done a better job myself

Indeed, I hit the stop button quite a bit when Claude goes off course. Then make a note of the right choice so maybe it won't do that again, revert and proceed. I have the feeling there is an optimal size of project proportional to the context size, where you can fit the critical design points into the context and/or there are enough examples in the code of how things should be done.

For me, one of the new superpowers is the ability to interactively do multiple drafts following different design principles and see which works better.

I just started an embedded project where two different people had implemented subsystems independently, and I asked Claude to merge the code into a single project and convert the existing synchronous code into asynchronous state machines called from a single main loop. It wrote three drafts with me giving it different stylistic principles to follow. I don't know if I would have had the patience to do that myself!

That’s always true in my experience, but that doesn’t necessarily mean you need to. The trick I’m working towards is refining the workflow such that i can reliably produce maybe 90% as “good” as what I’d personally produce but much, much faster. All sorts of side work I was avoiding before also becomes much easier, less tedious refactors and large test coverage and etc. It can type much faster than I can, the trick is if we can constrain the thinking enough to make it useful. Keeping it as an autocomplete is as productive as it is difficult imo.

presumably you've done the work enough pre-AI to notice those mistakes?

> I'm able to quickly stop it and fix any mistake it makes

I would think that's the process too, but according to the article the dude is almost completely hands off:

> You come back to ten thousand lines of code. You spend 5 minutes reading. One sentence of feedback. Another ten thousand lines appear while you're making lunch.

You can't humanly review 10 thousand lines of code in 5 minutes. This is either complete bullshit or it really writes flawless code for them and never makes any mistakes.