OpenAI Codex lead on the new shape of product work | Andrew Ambrosino

Key Concepts

• Codex: An OpenAI desktop application that has evolved from a developer tool into a general-purpose productivity platform.
• Taste: The critical human judgment required to curate, refine, and direct AI-generated outputs, moving beyond mere implementation.
• Agentic Workflow: A shift in product development where individuals act as "builders" who use AI to execute tasks, rather than relying on traditional, rigid product management processes.
• Role Collapse: The blurring of traditional boundaries between engineering, design, and product management, where an individual's role is defined by the "average" of their daily contributions.
• Zone Defense: A product management framework where team members spread out to cover gaps in a chaotic, high-velocity environment rather than working in overlapping, redundant clusters.
• Vibe Coding: A colloquial term for using AI to rapidly iterate on software by describing intent and steering the model, rather than writing code character-by-character.

---

1. The Evolution of Product Work

Andrew Ambersino, product and engineering lead for Codex, explains that the "shape" of product teams has inverted. Previously, implementation was the most expensive part of software development, necessitating heavy research, documentation (PRDs), and de-risking before building. Today, because AI makes implementation cheap and abundant, the bottleneck has shifted to taste and curation.

• The 90-Prototype Problem: Instead of writing a single document, teams now generate dozens of prototypes. The challenge is no longer "can we build this?" but "which of these 90 attempts is worth pursuing?"
• The Death of the PRD: While some argue documentation is dead, Ambersino argues for choosing the right medium for the goal. Documents are still useful for clarifying vague areas, while prototypes are essential for stress-testing interaction patterns.

2. The Role of "Taste" and Human Judgment

Ambersino defines "good taste" as a combination of aesthetic judgment, systems thinking, and strategic alignment.

• Beyond Aesthetics: Taste involves understanding how a feature fits into the broader system, knowing what to prioritize, and deciding how to present information to the user.
• Why AI Struggles with Design: AI is currently better at software engineering than design because code is easier to "grade" (e.g., does it compile?). Design requires a human feedback loop based on taste, novelty, and cultural context—elements that are harder to train into models.
• The Abstraction Layer: A significant hurdle for AI is the "deep" design layer—understanding how different components in a codebase relate to one another semantically, which is crucial for long-term maintenance and rebranding.

3. Organizational Structure: "Member of Technical Staff"

OpenAI utilizes a "Member of Technical Staff" (MTS) model, which minimizes rigid job titles.

• Role Fluidity: Roles are not fixed; they are the "average" of what an individual spends their time on. A designer might spend 40% of their time writing code, while an engineer might spend 30% on product strategy.
• The Danger of Eliminating Roles: Ambersino warns against companies blindly eliminating the "Product Manager" role. He argues that product management is a discipline with "knowable best practices" that should not be abandoned just because everyone can now write code.

4. Planning in a High-Velocity Environment

Planning at OpenAI is intentionally "hazy" beyond the short term.

• The Planning Framework: The shorter the timeframe, the more detail is required. Long-term plans (9+ months) are kept vague because "any amount of precision added to a 9-month plan right now is false precision."
• The "Bake" Strategy: If a feature is too ambitious for current model capabilities, the team prototypes it, lets it "sit and bake," and re-tests it whenever a new, more capable model is released.

5. Real-World Applications and "Computer Use"

The Codex app is increasingly acting as an "operating system" for work, connecting to external tools.

• The Premiere Pro Case Study: A videographer used Codex to edit videos by having the AI write an extension for Adobe Premiere Pro. The AI then communicated with the extension to perform edits, demonstrating how Codex can act as a bridge between disparate professional tools.
• Automation: Users can set up "personal systems" within Codex, such as filtering emails or managing Slack notifications. Ambersino emphasizes that the goal is to make Codex a "home base" where users start and end their work, using the AI to orchestrate other applications.

6. Synthesis and Conclusion

The core takeaway is that we are moving toward a future where agency is the primary currency. The most valuable individuals are those who can shepherd an idea from inception to completion, possessing the taste to distinguish signal from noise in a world of infinite AI-generated content.

Ambersino’s advice to those navigating this shift is to avoid getting married to a specific process. Instead, focus on the outcomes you are uniquely able to deliver and be willing to change your methodology as AI capabilities evolve. As he notes, "The implementation is actually not the expensive part anymore. It's dare I say taste."