Your Brand Standards Should Talk Back

Every B2B company past a certain size has brand guidelines. Almost none of them have brand guidelines that do anything. The PDF says "warm, human photography"; the actual images ship from whatever a designer, an agency, or lately an image model felt like on a Tuesday. The standard and the output drift apart, and the only feedback mechanism is someone senior saying "this doesn't feel like us" in a meeting where nobody writes down what to change.

In T2D3 OS, brand standards are not a document. They are direction cards — per-medium style rules (illustration, photography, iconography, graphic elements, plus logo, color, and type) that every generator in the product reads before it makes anything. A hero image, a blog illustration, a deck icon: each one is styled by the same few cards. Which means the cards are worth arguing with.

As of this week, you can.

Feedback goes where the standard lives

The change is small on the surface: Max, the assistant docked next to every screen, can now read your direction sets and revise them from plain feedback. You say "our photography direction should be warmer — candid people mid-conversation, never posed stock staging" and Max runs the same skill that wrote the direction in the first place, with your words as the highest-priority input.

What matters is everything that rides along under the hood:

The revision is grounded, not improvised. The generator doesn't just paraphrase your sentence into a rule. It re-derives the direction from the same evidence the original came from: your signal library — the uploads, call notes, and research you've fed the system — plus your locked brand identity, the style examples you've endorsed, and the running tally of every 👍/👎 your team has cast on generated assets. Your feedback steers; the evidence anchors. And each revision records exactly which signal sources grounded it, so the cards can say so.

Human judgment is load-bearing. Direction cards you wrote by hand, or pinned, are protected — a refine can add around them and rewrite the AI's own drafts, but it cannot silently overwrite a rule a person committed to. This is the same lock-gate philosophy the rest of the foundation uses: the AI drafts, the human decides, and decisions don't evaporate because someone typed a new sentence in a chat box.

It's the same door, not a new one. There is no separate "brand chat." The panel you already use for strategy questions and ICP edits is the one that tunes your illustration style. One assistant, standing next to the thing it's editing, updating the page in place as the conversation lands.

The moment the AI said "not so fast"

The best part of shipping this was watching the live acceptance test do something we didn't script.

Our test brand's photography direction had two cards, both human-written: one demanding a product render on pure black — a single upright product mockup, dramatic and empty — and one about warm dark staging. Then we sent the feedback above: warmer, candid people, no posed staging.

Max did not run the refine. It came back and said, in effect: your feedback partly contradicts what's pinned. "Candid people mid-conversation" sits awkwardly next to "a single product mockup on pure black" — that's a product-render rule, not a people rule. It laid out two paths: add a third card for candid people photography and leave both protected cards intact, or unpin the conflicting rule so it can be rewritten — and it recommended the first, with the reasoning attached. Then it waited.

That is the behavior we build for. Not an AI that does what you say, and not an AI that does what it wants — an AI that notices your instruction collides with your own earlier judgment, names the collision, and puts the decision back in your hands with a recommendation. We said "run it." It ran the refine, kept both protected cards untouched, added the candid-photography guidance around them, and the cards updated on screen next to the chat with an undo button on the tool card.

If you've read our post on devil's advocates, this is the same principle wearing different clothes: the assistant's job is to strengthen the human decision, not to replace it — and definitely not to bulldoze the decisions you already made.

Taste is trained by choices

The second half of this ship is smaller but, we suspect, more used. When you generate a single asset — say, one icon from a text brief — the result used to land somewhere in the asset library above, one silent file among many. Easy to miss; impossible to compare.

Now the generate box produces three variations, rendered right under the button you pressed. Each carries a thumbs-up/down. You pick one to keep; the pick itself is recorded as a positive vote, the losers are discarded — along with the vector twins the pipeline auto-derives, so nothing orphaned lingers in your library. And here's the detail we're most deliberate about: votes survive deletion. Thumbs-down a miss before you discard it and that judgment stays in the system, feeding the learned guidance that steers every future generation. The variations you reject teach the style as much as the one you keep.

This is the quiet thesis under both halves of this release: a brand isn't defined by a document, it's defined by an accumulation of choices — this image over that one, this rule sharpened after that miss. Most tools throw those choices away. We treat them as the most valuable signal a marketing system can collect, because they're the one input no model can generate: your taste.

Why this matters beyond aesthetics

Answer engines and AI-assisted buying are making brand consistency a distribution problem, not just a design problem. When more of your surface area is generated — pages, decks, social cards, icons — the standards that govern generation are your brand. A static PDF can't govern anything. A direction set that every generator reads, that your whole team can tune by talking to it, and that gets sharper with every thumbs-down? That's a brand standard doing its job.

The directions tab is live in the Visual Brand module for every T2D3 OS workspace. Open it, press ⌘J, and tell your brand standards what's been bothering you.