Taste is the whole job
AI made producing brand work nearly free, and that broke something quiet: point a model at a blank page and it hands back the average of everything it has already seen. The scarce thing was never the ability to generate options. It was the judgment about which option is yours.
Inside wild we kept hitting the same wall. A designer would spend a morning re-explaining the brand to the machine, get something good, and by the next session the machine had forgotten all of it. Taste does not survive being retyped into a box every morning. Judgment needs a machine-readable home.
Files, not platforms
That's what the Brand Context Protocol is: the one place that holds everything that makes a brand itself, its voice, its design, its rules, written as plain markdown and YAML in a git repository the brand owns. No build step, no vendor database, nothing that can't be read by a person or forked in a minute.
The structure has four parts. Brand Truth is the source: voice, design rules, beliefs, and, crucially, refusals: the things the brand will never do. Decisions are append-only and dated, so the reasoning is auditable instead of silently overwritten. Skills are reusable processes: a landing page, a deck, a campaign email, each one pulling from Brand Truth, each new one a permanent capability. Outputs are generated from that source rather than from blank prompts, and every output carries a trace showing how it was derived. And Brand Check scores the work against the standards, feeding what it learns back into the source.
The result is a loop: the more the brand ships, the sharper its own rules become. Evidence compounds into taste.
Context engineering, for brands
If you build with models daily, the shape is familiar. Brand Truth is curated context: the durable facts, rules and refusals a model needs in its window before it generates anything. Skills are versioned, reusable procedures the model loads for a task, the same pattern the frontier labs now use for their own assistants. Brand Check is an eval suite: it scores outputs against written standards and feeds the failures back into the source. BCP is context engineering applied to the asset most companies still keep in a slide deck: their judgment.
Why it sits where it sits
The protocol deliberately lives between the model and the tools. Models will keep changing; the brand shouldn't. Make the AI swappable and the brand constant, not the other way around, which is exactly what the proprietary "AI brand platforms" invert when they trap your identity in their database.
The format is open source. The honest admission is that the format was never the valuable part. What matters is what you put in it: the refusals, the taste, the accumulated judgment of people who know what the brand feels like when it's right. The machine makes the options. People make the calls. This is just the first system we've built where those calls stay made.