Token Graph

A token graph is the central spec that holds every brand rule as structured interconnected values the AI can read on first pass. Color nodes carry hex values semantic roles contrast floors approved contexts and forbidden pairings. Type nodes lock specific font stacks weight ranges line height multipliers and tracking rules. Voice nodes replace every vague adjective with measurable behaviors such as sentences under 18 words lead with the core claim no hedge words and zero filler openers. These nodes wire together. A change in base spacing token updates every related ratio across layouts motion and imagery. The graph sits at the core of the AI brand system. Prompt packs reference it directly. Evals score output against its exact values. The editor tunes the connections when new edge cases appear. Without this graph the AI reconstructs the brand from scratch on every call and drift becomes inevitable.

The token graph is not a upgraded Figma style guide. It is not color swatches with clever names or a Notion page of personality sliders. It is not the old brand book dressed up with variables. Those tools worked when one designer shipped one asset a week and applied taste by hand. They collapse at AI speed. Adjectives like warm but professional or bold yet approachable give the model nothing to lock onto so it defaults to training data biases. The 2024 Klarna campaign showed the failure mode with inconsistent color casts and proportions the team could not patch fast enough. Coca-Cola's Create Real Magic campaign in the same period flooded the internet with outputs that ranged from on-brand to completely off because their system offered only loose logo constraints. The token graph kills that gap by turning every subjective rule into data with no room for interpretation.

Vercel runs the clearest live token graph with Geist. The system powers v0 every marketing surface the documentation and the product itself from one connected source. Color tokens like foreground-primary link to background-surface tokens with hard contrast minimums of 7:1 for text. Those connect to typography tokens that enforce Geist weights between 400 and 700 with preset scale steps. Motion tokens define easing curves and durations that tie into accessibility nodes for reduced-motion preferences. Voice tokens specify sentence length under 20 words direct openings no corporate filler and strict ban on words like actually or essentially. When the AI generates campaign assets or UI variations the prompt pack pulls the full graph the eval layer scores every output for compliance and failures route straight to the editor. Linear applies the same rigor to voice tokens that govern release notes in-app copy and changelogs. Their graph includes paragraph length targets reading grade level and required sentence structures so AI writing tools stay in lockstep. Stripe treats layout grammar tokens as brand tokens. Their docs AI generates pages that match the exact grid ratios heading scales and rule weights used in handcrafted surfaces. Figma connects component variants directly to the brand graph so AI-generated plugin interfaces cannot introduce new proportions or color logic. Anthropic runs automated voice evals against their graph across all generated documentation. These companies do not police individual assets. They govern the graph and the system stays consistent at ten thousand outputs a day.

Roll out a token graph the day your AI volume exceeds what a human can review. That threshold hit most teams in 2024 and became table stakes by 2026. Use it for social variants email campaigns product visuals ad concepts and website copy. Pair the graph with scoped prompt packs that reference its nodes with evals that run contrast voice layout and motif checks and an editor who reviews aggregated failures instead of every piece. The system pays for itself the first time it catches drift that would have shipped publicly. Heinz got away with loose AI experiments in 2022 only because a century of ketchup imagery saturated the training data. Most brands lack that advantage and lose control without the graph.

Skip the token graph for low-volume human-driven work. A local cafe posting twice weekly needs basic swatches not a connected spec. Avoid it if your brand director still wants to hand-approve every output instead of shifting to system editor. Without that role change the tokens sit unused and the old PDF leaks return. Run the AI-readiness checklist. Fail on token coverage or eval loops and you are operating a 2018 system in a 2026 environment.

Token graphs replace subjective brand books with connected machine readable values so AI can scale your brand without turning it into visual noise.