There is danger in over-fitting the agent. Fast release cycles require frequent model switching, sometimes to another provider altogether.
In the end, it's hard to measure improvement. Most of the advice out there is vibes based, treat it as such.
You cannot write the files directly nor can you download them directly, but you can ask a chat to show you the content or only the mermaid diagram that exists in workspace.md (or however you want to name it), and mermaid can represent quite complex networks out of the box, their relations, what is blocking something else, and so on. Chats can then edit this file with search-and-replace instead of re-writing it all over again.
For those that didn't read the article, one of the first things it says, in bold nonetheless, is "I didn’t design this system. I arrived at it"
(I am a PM of 15 years trying to navigate all of this…)
jdw64•50m ago
The code that goes into PLC equipmentss and hardware is incredibly simple—but it's still a product. If you look under the hood of a lot of software, the code is a mess. Still, it's a product. And when Claude's code leaked, everyone laughed at the code hygiene—but a lot of people still used it.
What is a product, really?
tolugenius•33m ago
orochimaaru•32m ago
There is a lot of hand wringing in the software space about ai code being a mess - but from a product perspective code quality doesn’t matter. It’s the reliability of what people (or agents) use that matters.
Code quality used to further that goal but only because it was really hard to rewrite or fix code. If that barrier is removed does anyone really care what the code quality under the hood is?