This is one reason I’ve been anchoring the other end of my thinking distillation column with paper notebooks. I’m filling them up at like 10x my rate from before AI (XL Moleskine fills up in 3 months instead of 2-3 years these days). LLMs are most naturally one-shotting tools. They’re best as either one-shot probe, or a series of such probes (think exploratory well-digging) to map the distribution of idea “resources” in a geography OR for drilling a production well once you’ve decided where to drill.

The “orientation and sensemaking” part, between discovery/exploration and one-shotting the execution so you can then switch to finishing touches, is best done not just outside of LLMs but outside a computer entirely. This is because even computer tools are destined to have AI seep in. Roam has an AI integration I trued to use then turned off. Obsidian I believe would let you run with Claude cowork in the directory. Such uses would turn these notebooking tools into either probing or execution tools.

Ie orient and make the big decisions in a paper notebook, do all the operational observe/decide/act with an LLM. It’s a paper-interrupted OODA loop for me.

These days an idea doesn’t exit my notebook and enter a digital workflow until after I have an orientation lock on it. On the other end, probe/discovery don’t enter my paper notebooks until I have enough situation awareness to make little sketch-maps by hand without referring to my LLM chats.

The notebook is nice because it’s capped at pure human capacity. I do sometimes interrupt a notebook session to do an LLM probe or make a few digital notes in Roam. And occasionally, I’ll put a pointer to an LLM chat or Roam page in my paper notebook, where there are details that matter even if the pointer is all I need for orientation thinking.