My Approach to AI

Most designers using AI in 2026 are moving fast. The failure mode isn't slowness. It's confident wrongness: a tool that produces plausible output while being wrong in ways that are hard to catch.

The method has two layers. The first is a wiki: every project, every decision, every source, logged, cross-referenced, filed. Only ingest, never invent. It holds the context that gets lost six months later, when the original reasoning is gone.

The second is critique. Before anything ships, I run the work through twelve models in parallel. Each gets the relevant wiki context first. Each has a different role: one checks substance and logic, one flags tone, one cuts redundancy, one checks sentence mechanics. When two different model families flag the same issue, I fix it. When only one flags it, I decide. No model generates anything.

I built this for design work, then deliberately tested it on completely different problems: a skincare app I coded, reorganising my wardrobe, planning my kitchen layout. Not because those problems needed this level of rigour, but because I wanted to know if the method held when the stakes were low and the domain was unfamiliar. It does.