alt Hacker NewsSame. I do not use LLMs for writing and I recommend everyone i work with to do the same. Even for code, I come to the code with the idea of what I want and provide that to the agent, then we iterate on what needs iteration or decisions.
Just did this for some caching, started with the structure of what I wanted the cache to look like, asked the agent to start the work, didn't like how the architecture came about, scratched it and rewrote the whole thing, so it fit the model I now wanted.
Was also just having this discussion with friends, that I can only think seriously about a subject if i can put it to paper (even virtual paper). Writing lets me organize my thoughts, clearly define my assumptions and see if any of it make any sense. I can't imagine what it would be like if i couldn't write, my brain just doesn't work without it.
I now have dozens of projects where I've embraced the yolo. I went through elaborate systems, workflows, review triages, architectural linters, specialist agents, yet, they all still suffered the same fate - slop which I don't understand and now LLMs don't understand either.
The only thing that worked was a standing instruction and periodic system reminder injected from the harness:
"If any assumption doesn't hold, if there's a fork in the road, any architectural decision needs be made, STOP and report back to the user. Do not try to push through the problem."
This has worked remarkably well for me. Now I have to think a whole lot more. It's much slower, yes, but I don't really see any other way that doesn't end up in garbage.