Building AIR's brand and site — with AIR

There is a fair test for a framework that claims to bring structure to AI work: use it to do something real, then report honestly on what happened. So the brand you are looking at — the palette, the mark, the brand book, the templates — and this website were all built with AIR. This is an account of how that went, including the parts where the structure mattered and the parts where it couldn't help.

The setup

AIR — short for AI Resource — is a prompt-based framework. It has no model of its own; it runs on whichever chatbot you bring. For this project, AIR ran on a frontier model — Anthropic's Claude (Opus) — while I steered. AIR supplied the structure: it scoped the work through onboarding, kept a roadmap, gated each step for approval, printed its reasoning as AIR objects, and carried the project across sessions with handoff cards. The model did the generating. I made the decisions and caught the mistakes.

What got built

Across a series of sessions, the work moved from a blank brief to a complete system: positioning and voice, the Ember palette, the type pairing, the dot-in-frame mark, a sixteen-section brand book, a context pack, working templates for decks, documents, email, and a one-page reference — and then this multi-page site. No single session could hold all of that. The project stayed coherent only because each session ended with a handoff card and the next one booted from it.

Where the structure mattered

Three things earned their place.

• Onboarding stopped the drift. Before anything was produced, AIR fixed what we were building and how it should work with me. That agreement scoped every later step.
• Gates kept it honest. Nothing shipped without passing a checkpoint and my approval. When a diagram's labels were ambiguous or a color felt off, the gate is where it got caught and fixed — not after the fact.
• Handoff cards made a long project possible. A brand-and-site build doesn't fit in one context window. The card carried the state forward, session after session, so the work resumed exactly where it left off.

Where it couldn't help

AIR didn't make the output correct, and it doesn't claim to. The quality of every draft depended on the model, and the model made mistakes — a misaligned diagram here, an overlapping element there, a placeholder left in. AIR surfaced the work for review; I was the one who had to notice. Strip out the human judgment and the structure still holds its shape, but the results wouldn't. That is the honest shape of it: AIR makes the loop reliable, not the answer guaranteed.

The takeaway

What structure bought, in the end, was traceability and continuity — a creative and technical project that ran across many sessions and stayed coherent, auditable, and resumable. It didn't replace the work of deciding and checking. If anything, it made that work easier to do well.

You can read the whole brand system it produced, or boot a session and put it to the same test yourself.