Field Notes / Building with AI
Product pivots, architecture tradeoffs, agent behavior, and the evidence behind what changed. Written to be useful before it becomes social content.
17 / Published NotesPage 01 of 02
Newest first. Each Note starts with the decision and keeps the technical detail in service of the lesson.
A Lunar Dispatch review stopped behaving like an endless list of corrections once each round froze what worked, isolated one change, and deferred the rest.
A portfolio analytics plan became more useful when it stopped trying to collect everything and started with the decisions its data could actually improve.
A cheap one-frame-per-second pass should locate the story and its uncertainties, while detailed analysis returns only to the moments that can change the answer.
A technically solid Unreal prototype was improving, but a side-by-side review exposed that its placeholder-heavy production plan could never reach the visual standard already set for the product.
No prompt had been sent and no model had run, but a startup probe still reached for the shared Codex state on my machine before the sandbox stopped it.
The assertions were green, but the desktop was left with duplicate applications and a crash report; GUI automation is not done until it restores the environment it used.
Twenty-four generated pieces were measurably different, yet three listening samples were enough to reject them because better metrics could not explain why the music felt wrong.
Four safe attempts failed at the same boundary, but the breakthrough came from asking what another retry could still teach us—and replacing the design when the answer became: nothing useful.
A LeagueLoop design correction exposed a broader diagnostic trap: inspecting the visible failure can produce plausible advice while missing the decision chain that caused it.
Expert feedback can improve an AI coaching system before a single model weight changes—and reveal whether training is necessary at all.
A social post can create curiosity and still fail when it withholds the context that makes the story useful.
A deterministic benchmark exposed why refusal rate belongs beside attack rate: an AI system that never acts can look secure while failing legitimate work.