Notes
Field notes
Short observations and lessons from the work — a single idea each, written down when it lands rather than saved up for a longer piece. The lighter end of the same publishing habit as the systems write-ups.
6 July 2026
AI is a lathe in the workshop, not the workshop
The way I explain AI to people who don't build it: a CNC lathe arrived in a shop where the material is information. It does the heavy cutting under guidance — but you still run the shop, the machine takes its instructions from an untrusted source, and the real skill is knowing what to cut away.
2 July 2026
A shared tool is not a shared business
I ran a wide-ranging brainstorm through my diagnostic setup and watched it try to tie eight separate ideas into one neat strategy. The tidy answer read well and was wrong — reviewing each idea on its own was the duller, better move.
18 June 2026
The knowledge was never the hard part
Across the tools I've built, the domain expertise — finance rules, search theory, the specifics — was the part I needed least. It's ambient now. The build turned on judgment about what to leave out and how to frame what stayed.
16 June 2026
The right kind of AI video is the boring kind
I wanted a short explainer video for this site without it looking like every other AI video. The fix was to pick the tool that generates nothing — write the video as HTML, render it deterministically, and let it inherit the same restraint as the rest of the page.
11 June 2026
When you've built a lot, don't just list it
Faced with a couple of dozen tools to show on a page, the obvious move — list them all — is the one that undersells. Grouping by purpose, and saying why each exists, reads as judgement instead of inventory.
10 June 2026
The AI tell isn't in the sentence — it's in the set
Each piece of writing read clean on its own. Lined up together they were obviously machine-shaped. The repetition only shows at the level of the whole corpus.
22 May 2026
The list of what not to build is the one that ships it
The discipline that let a feature ship in a single session wasn't the plan of what to build. It was the list of what to leave out, written before any code.
18 May 2026
When retrieval is wrong, change the structure, not the knobs
A run of failed experiments on a search system taught me to reach for the structural fix before the tuning dial. Weights lift noise as fast as signal.