11,000 lines of code on a new app in four days with two windsurf instances and void running continuously.
This is exciting... but it presents clear hurdles to QA and testing.
In these situations unit and flow tests become more important than they've ever been. I've been doing them at the end of every layer of the project to test milestones.
I'm trying to keep my test surface manageable between milestones.
And, by the same token, I'm trying to catch all the little details that get lost between context switches.
Punchlists, roadmaps, multi-agent QA audits and consensus building helps.
Another problem I'm having is useful, thoughtfully designed features that make sense, but that I never asked for. Especially on this new DeepSeek model. "But wait, what we really need here is a cloud based security and authentication layer!" and then it spends six hours implementing oAuth while I'm sleeping.
While some creative thinking is actually useful, and what I'm going for. I'm finding that some of these models are losing track of the scope of any given application. Which is why you need to say things on your task lists like, "This api is designed for desktop use and should be self contained."
But, even still with all the drawbacks, and there aren't many if you're well organized and thoughtful in your design process, it's a win.
Next step... removing the human from the loop altogether.
More on that later.