Logbook 2026-11-03

Today a quote turned into a way of working.

In the evening I read Katie Parrott’s How I Escaped AI Autopilot, published the same day on Every, and kept a quote from it in the Feed, about the moments when we should slow down. Her point is that polished output looks finished, and we easily take it for correct. Little by little, the AI takes the judgement calls, not just the labour. To keep them, she makes three changes, and the first one is to think before you look: write down your own position before asking the model anything. What I think, what I know, what I’m unsure about, what I refuse to say, what would make the piece useful.

That first change covered what I needed first: a real process of exchange between the AI and me, almost maieutic, where what I think comes out of me. The same evening it became a skill for this site, writing-prep. Before any article or essay, Claude takes me through her five questions, one at a time. It asks, then waits. It pushes back when an answer is vague, and keeps my own words in a brief. The draft comes later, and only from that brief. I simply thought it would be a good idea.

Deep down, I feel the risk is high, and I want guardrails. With a thousand ideas a second, and the means to build them now that AI keeps pushing back what’s possible, I want to go fast. But I have to know when to slow down, at the key moments, to make the right choices and not hand too much authority to the AI. It can so easily make things feel safe, or aligned with me.

The same day, I also read Matt Ström-Awn’s Code review was never about the code, which makes a close point about code. Code review was meant to check that the software does what we want, and we ended up asking whether the code looks good. The same trap: answering the easy question instead of the real one.