Golden Book MCP

I built Golden Book because I kept running into the same problem while working on FoldOps and developing website concepts for local businesses. Coding agents could produce functional frontend code, but the visual decisions often collapsed into the same generic layouts, shallow interaction patterns, and invented marketing claims. I wanted a way to give a builder stronger design judgment without forcing every project into one template.

Golden Book is an MCP design router and packet compiler. It reads the actual brief, identifies a compatible design family, chooses one primary golden example to own the composition, and retrieves additional patterns only when they solve a real gap. The packet also carries responsive behavior, accessibility expectations, interaction states, performance guidance, anti-copy rules, claim-realism constraints, source provenance, and browser QA gates.

Golden Book is especially useful for smaller, local, and open-weight coding models that benefit from richer implementation context, while remaining useful across model tiers because it improves what design evidence reaches the builder rather than replacing the builder itself.

I used Codex with GPT-5.6 to refine the routing system and the golden sets used by the router. It helped me inspect weak generated interfaces, compare original and routed builds, find cases where sources were too compact or too similar, and implement improvements to retrieval, source hygiene, optional patterns, and packet completeness. I made the product and design decisions, reviewed the generated work visually, and kept iterating until the router produced context I would actually trust in my own frontend work.

The hardest part was improving design context without creating a copying system. A source can be visually useful while carrying the wrong identity, unsupported claims, or an interaction that does not belong in the target. Golden Book addresses that by giving one anchor ownership of the composition, gating incompatible sources, tracking provenance, and treating optional patterns as focused mechanics rather than pieces to paste blindly.

The featured demo is a controlled comparison. The same OpenCode MiMo builder received the same martial-arts gym brief in both conditions. The second condition also received a Golden Book packet. Both pages passed the same browser checks at desktop and two mobile widths. The comparison does not claim that routing wins every time; it shows how better design context changes the decisions available to the same builder.

Golden Book began with websites, but the architecture is not limited to web marketing. I am building the library to support operational apps, dashboards, creative tools, game menus and HUDs, commerce flows, and other interface families. The router can keep expanding without changing the builder model or flattening every project into the same visual language.

The biggest thing I learned is that frontend quality is not only a model-capability problem. It is also a context-selection problem. More prompt prose is not a substitute for a coherent reference, detailed implementation evidence, honest constraints, and rendered browser review. My next step is to keep improving the golden sets, routing judgments, visual evidence, and evaluation loop so that the system becomes useful across more kinds of products.

Built With

• context
• mimo
• model
• v2.5