Programming

I worked on desloppifying a codebase, but had to do with AI. This was a job. Spec driven development didn't work on that project but targeted refactors did.

Setting up a project to work with AI is not easy, IMO. You need rules, and skills, and commands, and whatever else they call it in order to be productive, otherwise you spend your time fighting with the the AI.

You're right to question 10x and 100x productivity gains. Those only exist if you're not willing to look at the code. I've had people tell me "code isn't important", then run into issues, and ask me to help out. The code was indeed important. Therefore, how you use AI is the important part here.

Using AI to write the entire codebase is cute. It might get your project to MVP very quickly and then you have to hire somebody who knows what they're doing to fix all the stuff in it. I would recommend using it in a targeted manner:

• ideation
• code analysis
• targeted refactors
• boilerplate
• mini features
• bug detection

As an example, connecting an AI to an MCP like Ghidra for working with decompiled binaries is impressive. You do have to stay on your toes all the time and verify what it says, but it changes the while experience.

Once you let it loose to make bigger changes, it will inevitably fuck up while spitting out reams of code at you. Reviewing all of that is the bottleneck because it requires Actual Intelligence. Using another Artificial I to review what the previous AI did is only useful for bug detection or exploit detection (in my experience). For architectural mistakes or clean code (whatever your perception of that is) has only lead to pain for me.

So yeah, not 100x, not 10x, but maybe 1.1-1.5x. More if you're careless or don't know what you're doing and Just Want It To Work ™️.