Codex H9000

Give your AI a presence—one brief becomes a stateful visual identity directed by GPT-5.6.

Comment

Inspiration

AI products have increasingly capable voices, tools, and reasoning, but most still appear as a generic sparkle or spinner. Builders need a visual system that can communicate what an agent is doing without inventing a motion language from scratch.

Codex H9000 turns the character of an intelligence into a reusable, stateful visual presence.

What it does

A builder describes how an agent should feel, then selects a visual language and mood. GPT-5.6 directs a complete identity—name, six-color palette, geometry, motion, and behaviors for idle, thinking, speaking, and tool-use states—through a strict JSON schema.

The result renders immediately, can be presented full screen, saved to an identity library, or exported as a JSON contract for another product.

The interface also has an explicit deterministic Demo mode. Judges can test the complete flow without an API key, while the product clearly distinguishes local output from live OpenAI output.

How we built it

Codex helped audit a much broader pre-existing HAL Builder 9001 prototype and rebuild its strongest idea as a focused submission. The clean edition uses React, TypeScript, Vite, and a small Node server.

At runtime, the server calls the OpenAI Responses API with GPT-5.6, low reasoning effort, store: false, and strict structured output. A JSON schema constrains color roles, geometry limits, motion ranges, and four behavioral descriptions. The server validates input, caps request size, clamps render values, and keeps the API key out of the browser.

Codex performed the repository audit, product reframing, clean application rebuild, OpenAI integration, interface polish, contract tests, responsive browser QA, submission write-up, and video assembly.

How we used GPT-5.6

GPT-5.6 is the runtime visual-systems director. It translates subjective personality language into a predictable, machine-readable design contract. Strict structured output makes that creative direction safe to render immediately instead of asking the client to parse free-form prose.

Challenges

The hardest problem was preserving creative range without allowing model output to break the renderer. The solution is a deliberately bounded schema plus server normalization.

A second challenge was making the app honest and judgeable without configured credentials. Demo mode solves that while keeping provenance visible in the interface and exported data.

Accomplishments

  • One brief becomes a complete visual system, not a single image
  • The same identity communicates four meaningful agent states
  • The OpenAI key never enters client code
  • The complete flow remains testable without credentials
  • Full-screen presentation, identity library, and JSON export
  • Local type, lint, contract-test, production-build, dependency, and browser QA

What we learned

Structured output can bridge subjective creative direction and deterministic product behavior. The model does not need to return UI markup; it can direct a safe design system that the product renders consistently.

What's next

Next steps are an embeddable package for agent apps, streaming state events, reusable identity versioning, and direct design-token export for product teams.

Built With

Share this project:

Updates