I've been guilty of this and gotten pushback from my manager: "this feels like homework, cut these options down to 100 words each, max".
Curation and refinement are even more important when you can have genAI generate reams of text.
Seeking outside signals is even more important, like talking to customers, looking at real usage data, and more. It's too easy to trust believe what Claude tells you, even if you say "please argue against this idea", which you always should.
The word "more" there is doing a lot of work.
What is the "more"? Is it:
- more documents and text or more understanding
- more code or more valuable features
- more things to throw against the wall or more considered experiments
It's way easier to do the first things instead of the second.
Adding is always easier than reducing to the essence.
I find it interesting how herding agents has so much in common with being a team lead. Constant struggle between too detailed and too loose instructions. One difference though is that the team learns from you, but with agents it's only you who adapts. Saying that, because I don't count instructions or anything in the context window as adapting.
Someone will compete and take your place. Now you're expected to produce more, or step aside for someone who will.
Who decides value? I'm sure you can figure that out.
It matches the pattern of LLMs being very good at simulating the form of work output, which is an issue with code but it seems quite exacerbated with anything non-verifiable, like written communication.
I'm using Claude to write large files too, but it's a very iterative process and involves a lot of reading and correcting.
to be fair, i've been guilty of this with code. Ask claude to generate a python script that takes X as input and produces Y as output, run it, pipe to more, output looks ok but i don't check everything, write it to a file, send it on.