2026-07-25 · saturday
Assume I don’t know what I don’t know
This Saturday I sat down to do some overdue housekeeping on my dev setup, and instead of poking at config files myself I typed this into Claude Code:
This was inspired by Karpathy’s post on using voice mode to ramble on.
Interview me on my needs and my setup, help me ramble about the state of my workflow and find gaps and opportunities. Be brief, practical, pragmatic, avoid complexity traps and productivity hacks, assume I don’t know what I don’t know.
The last line was especially important as it tells the model to challenge my knowledge gaps and give me opportunities to learn and grow without over-engineering things.
That said, at some point it decided my laptop should run a local LLM for summarizing meetings, pulled a 5GB model, and only at runtime it discovered my machine has 16GB of RAM and chronically lives about 7GB deep in swap so I had to interrupt and steer the session:
you’re making assumptions such as “we need a model, go test model” — that’s a little optimistic and not an evidence-based approach that has awareness of my environment
✓ memory updated · profile-before-prescribing.md — measure the machine and the real concurrent workload first; let the budget math pick the tool
After that it profiled the machine, redid the budget for what’s actually resident during a call (e.g. the browser tab running the meeting, the terminal jobs, the transcription models, etc) and picked a smaller model the numbers could defend.
Pushing back also should make the AI permanently better at working with you, if you are often steering it then consider optimizing your prompt or your harness. (It’s tempting to just plow through, but it’s worth it)
“The smoking gun” as claude would have written this
The gaps worth finding are rarely in the config of your harness, they live in whatever you stopped questioning.
The result of this session allowed me to stop assuming my workflow couldn’t improve further and it automated more of my routine towards what I could delegate to new background agents, I highly recommend challenging your setup every once in a while.