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

What It Covers

Andrew Ambrosino, product and engineering lead for OpenAI's Codex desktop app, describes how AI has inverted the traditional product development process. With 5M+ weekly active users and 90% of all OpenAI employees using Codex weekly, he outlines how implementation cost collapse has made taste, curation, and judgment the new scarce resources in product work.

Key Questions Answered

• •Process Inversion: Traditional product development de-risked expensive implementation through documents and research first. Now implementation is nearly free, so the scarce resource has flipped entirely. Teams at OpenAI routinely generate 90 parallel prototypes of a single feature simultaneously. The critical skill becomes curating those explorations — identifying what works, what to combine, and what to discard — rather than producing the artifact itself.
• •Medium Selection Over Medium Elimination: PRDs are not dead, nor are prototypes universally superior. The correct approach is matching the medium to the specific goal: use a document when the problem requires clarity around a vague strategic area, and use a prototype when stress-testing an interaction pattern. Choosing the wrong first medium creates a "primal mark" that anchors all subsequent decisions to the wrong starting point.
• •Prototype Maturity Mismatch: AI-generated prototypes now look production-ready at the earliest exploration stage, creating organizational confusion. When a prototype visually resembles a shippable product, stakeholders pressure teams to release before the underlying research, user validation, or business logic is sound. Teams must explicitly communicate which stage of the design process an artifact represents, regardless of how polished it appears.
• •Role Definition by Average, Not Boundary: Functional roles at Codex are defined by the average of where someone spends time, not by hard boundaries between design, engineering, and product. Designers on the team write code; the PM holds a computer science master's degree. Eliminating role boundaries is valuable, but eliminating the concept of specialization entirely is dangerous — each discipline carries accumulated best practices that cannot be rebuilt quickly.
• •Zone Defense Product Coverage: With implementation abundant and ideas emerging from every direction, product managers should operate like zone defense in basketball — spreading across the organization to cover gaps rather than clustering together. The goal is ensuring every area of active development has a taste-maker providing steering, framing, and coherence, rather than concentrating product oversight on a small set of pre-planned initiatives.
• •Model Timing Determines Product Fate: The Codex app released in February 2024 would have failed if launched in November 2023 — the only variable was model capability, not product design. Teams should build features speculatively against future model capability, label them as not-yet-ready artifacts, and retest each time a model leap occurs. The shape of a feature can be correct while the intelligence required to execute it simply does not yet exist.