Instead of designing for current AI limitations, the team projects model improvements six months forward and builds interfaces for that future state. This framework guided the creation of their agent-first products, assuming capabilities would eventually catch up to the design.
For generative UI, designers move from defining exact layouts to creating a "skill" that guides the AI. This skill defines a design language, references tokens, and suggests best practices, steering the model's output without over-constraining it.
Instead of assigning designers to narrow features (e.g., "code review"), teams are structured around broader personas (e.g., "developer"). This ensures designers consider the entire user lifecycle and avoid optimizing for a specific metric at the expense of the overall experience.
Designers at OpenAI who have never coded before are now using tools like Codex to create branches of the main ChatGPT application to prototype ideas. This shift dramatically lowers the barrier to high-fidelity, native prototyping and changes how designers build and test concepts.
While tools like Codex empower designers to code, the most critical skill remains communication. The primary role of design is to create a compelling vision and bring partners along on the journey. A powerful prototype is useless if it can't persuade others to build it.
The design process has shifted through three phases: 1) Figma mocks, 2) isolated code prototypes, and now 3) "vibing in prod"—creating a branch of the actual product to test wild ideas in a realistic, high-fidelity environment.
OpenAI controversially designed Codex with the agent chat as the primary interface, relegating the code view. This "agent-first" approach, based on future model capabilities, was initially polarizing but became the standard pattern as models improved, proving the value of designing ahead of the curve.
Given the potential global impact of their products, OpenAI screens heavily for humility and genuine belief in the company's mission. They look for designers who are obsessed with user impact and approach the work with a sense of responsibility, not just as a "hot" career move.
To improve generative UI, designers create a "golden set" of 10-20 ideal prompts and their desired outcomes. They repeatedly run the model against this set, identify failures, tweak the guiding "skill," and iterate, creating a flywheel for quality improvement.
Overly prescriptive fixes or guardrails for current AI weaknesses are temporary. As models improve, this "scaffolding" becomes obsolete. Teams must stay flexible and be ready to remove old constraints with each new model release, rather than over-engineering for today's problems.