The 'Six Months Out' Framework
Ed Bayes advocates for a design philosophy that ignores current model limitations in favor of projecting where capabilities will be in six months. By building for that future state, the team avoids the trap of designing only for today's constraints. This approach allowed the team to build an 'agent-first' product for Codex, which initially felt controversial because it diverged from the standard IDE paradigm, but eventually became the standard as model capabilities caught up to the vision.
Evolution of Prototyping: From Figma to 'Vibing in Prod'
Prototyping at OpenAI has evolved through three distinct phases:
- Figma-based: Traditional static design and interaction testing.
- Code-based: Designers with engineering backgrounds building fake versions of the product to test interactions using live APIs.
- Vibing in Prod: The current state where designers create experimental branches in the actual codebase. These prototypes are not intended for production merge but allow for testing 'insane' ideas in a live environment, providing a high-fidelity experience that static tools cannot replicate.
Structuring Teams Around Personas
Bayes structures his design team around user personas rather than specific features. This ensures that designers maintain a holistic view of the user journey—from onboarding to power-user workflows. By owning the entire lifecycle for a persona (e.g., developers vs. non-technical knowledge workers), designers avoid the 'last-mile polish' trap and participate in early product strategy and exploration.
The Shift to Universal Interfaces
OpenAI is moving toward a universal interface for getting work done. While developers were early adopters of agentic workflows (like file editing and command execution), the goal is to make these powerful capabilities accessible to non-technical users. This requires a delicate balance of safety and agency: building sandboxes where users can approve agent actions (like sending emails or running code) on a case-by-case basis, ensuring the user remains in control despite the high level of automation.
Eval-Driven Design
Design at OpenAI is deeply integrated with research and evaluation. Designers work closely with researchers to understand the model's 'confidence thresholds.' When a model reaches a certain level of reliability, the interface can be simplified and abstracted. The design team uses quantitative and qualitative feedback loops—including weekly 'good, bad, and ugly' sessions—to stay grounded in real user behavior rather than designing in a vacuum.