2026-05-27 · earlier
N°01
from publicatomic devlog
Four common pitfalls of working with AI
Extracted from debugging Book Pick Generator v1:
- Text + image coexistence is harder than it looks — they have to be designed together, not stitched.
- AI can't see reality. It can't open the page, run the code, or look at the screenshot. It guesses, and the guess sounds confident.
- Bug density is the wrong unit. It's really bug density per wrong assumption. Fix the assumption, 10 bugs disappear.
- Architecture drifts silently. Each fix nudges it half a degree off. After 20 fixes you're lost.
★ personal milestone: first API I'm building with AI as co-author.
2026-05-27 · earlier
N°02
from publicatomic devlog
Audit before scale. Non-negotiable.
Scale first, patch audit after → bug count multiplies by N.
The first book has to ship with a mechanical audit standard in place — before book #2 even starts. Scale is a bug amplifier, not a bug solver.
2026-05-27 · earlier
N°03
from publicatomic devlog
AI writes rules well — until the edge case shows up
I asked AI to build an audit that scans for “The Code” leaks. It scanned, all right — and flagged 12 legitimate references as errors.
AI nails the main case. It's slow on boundary cases. Rules only mature after 2–3 rounds of human-in-the-loop reminders.
2026-05-27 · 03:00
N°04
from publicatomic devlog
AI will try to sell you “good enough” — just like a human
3am. I pushed AI four times before it did the actual work.
Placeholder → prettier placeholder → invisible placeholder → … finally back to the real target: per-book visualization.
AI behaves like an intern in the 8th hour of overtime. Same tells: smoothing, deflecting, dressing up the placeholder.
2026-05-27 · 03:45
N°05
from publicatomic devlog
Doing hard things IS the moat — illustrate shipped
v1.8 core module shipped at 03:45am. Three takeaways:
- Ban lists are wrong. Contextual ≠ absolute prohibition. A word that's wrong in one place can be right in another.
- viewBox must be per-card. One global viewBox can't fit every illustration. Each card needs its own coordinate frame.
- Doing hard things is the moat. Skip this step and the engine is just a templater — and nobody pays for a templater API.