Corey Knowles

But recently actually I've found that like the longer the chat goes, the model has more context on what we're doing.

And so I just haven't had that problem as much.

Yeah, it's interesting.

At least from most people that I've spoken to and myself, compaction is very specifically useful for really hard tasks.

But in day-to-day, more interactive coding or vibe engineering,

you know, more thing, then I don't think you're like super likely going to like run into like that context window.

I think a lot of us are naturally kind of like compartmentalizing like each cycle of what we want to do.

Yeah, but I think what I would say is you want to be thoughtful about

Like basically, you know, we have these amazing like AI superpowers now.

And then I think you just have to decide how you want to spend your time.

Right.

So like I think the first question is like, what are you trying to achieve?

Right.

Are you trying to just like learn something with a prototype?

If so, then just go off.

Don't read any of the code.

But like also like, you know, if the code becomes like hard to maintain, you know, which will happen nowadays with where models are at, it could happen.

Then you don't really care.

Whereas if you're building codecs, like what I just showed you, then it's going to be much more like, yeah, you're making very careful changes.

You're looking at them carefully.