Field notes
An open notebook on marketing × AI.
About marketing, about AI, sometimes both. I share what I try, what works, what doesn't, and what I'm still figuring out along the way.
Everything local: the setup that gives AI enough context to be useful
AI made the work faster, then context became the bottleneck. My current setup keeps project state, agent rules, machine-level tools, voice transcripts, and memory on disk so any model can start closer to the work instead of making me re-explain it every session.
MAY 11, 2026 · 9 minHandy + local LLMs: the dictation setup that quietly built a corpus
I started with a simple goal: type less. The useful surprise was the corpus left behind: 375 clips, 192 minutes of audio, raw transcripts, and a smaller set of cleaned pairs, all local and inspectable.
MAY 11, 2026 · 8 minThe double-tap: how a non-developer ships code worth trusting
I am not a developer, so I need a cheap way to catch confident technical mistakes before they become shipped work. My current answer is the double-tap: when a plan or diff matters, I hand it to a different model with different priors, then paste the smallest useful correction back to the writer.
MAR 15, 2026 · 6 minClaude Code vs Cursor? You don't have to choose, use both.
I was spending expensive model time on cheap implementation work. The experiment: keep Claude on scoping and judging, let a faster headless builder do bounded edits, then make Claude read the diff. It works for clear slices. It fails when the work is still being shaped.
FEB 26, 2026 · 7 min