How Cheap Code Flips Product Processes, Design, and Roles – Insights from OpenAI Codex Lead

This article distills an interview with Andrew Ambrosino, product and engineering lead for OpenAI Codex, highlighting how the cheapening of code implementation overturns traditional product workflows, design logic, and team role definitions.

1. From Expensive Implementation to Expensive Selection

Historically, high implementation cost forced teams to mitigate risk before coding—research, documentation, prototyping, and design reviews. Now, anyone can quickly prototype with AI, making the hardest problem deciding which of the many possible ideas is worth pursuing . The most scarce resource is taste and the ability to filter , not raw coding ability.

2. "PRD Is Dead" – Only Half True

The real issue is treating fixed steps and formats as the methodology itself. Teams must choose the right medium for the problem: documents excel at clarifying vague, boundary‑less product questions, while prototypes are better for validating interaction patterns. As production becomes cheap, choosing the correct medium becomes crucial because the wrong one can mislead.

3. "Looks Like a Finished Product" No Longer Signals Readiness

High‑fidelity prototypes may still be early‑stage experiments. Teams need to distinguish visual appearance from actual process stage to avoid anchoring on premature signals.

4. Design Process Still Exists, Rigid Steps Do Not

The traditional design pipeline relied on expensive implementation to justify thorough upfront exploration. Tools like Figma moved prototypes earlier; AI now pushes near‑complete implementations even before development. Teams must retain a higher‑level framework, asking:

Are we exploring, validating, or converging?

Is the artifact for clarification or for user testing?

Is it a research tool or a release‑ready version?

In other words, the rigid steps have died, not the awareness of which stage a work item is in.

5. Taste as Direction‑Judgment, Not Just Aesthetics

Taste encompasses evaluating worth, choosing form, integrating into the system, aligning with direction, and presenting value. When anyone can quickly build anything, the critical question becomes "What is the goal? Where are we heading?" .

6. Why AI Still Lags Behind in Design

Design evaluation involves subjective human judgment, cultural context, and novelty, making it harder for models to give binary feedback. Moreover, models excel at adding code but struggle with deletion, refactoring, and abstract design reasoning.

7. Feature Success May Require Multiple Releases

A feature can fail several times not because it is fundamentally bad, but because the underlying model at launch wasn’t smart enough. Success depends on the model’s capability at release time , not on the feature itself.

8. Long‑Term Planning Should Remain Vague

Short‑term tasks need concrete detail; long‑term plans should stay fuzzy to avoid false precision. Over‑granular long‑term roadmaps often lead to errors.

9. Role Boundaries Blur, Expertise Remains Essential

People are now defined by the average of the work they do rather than strict functional titles. Overlap increases—designers code, PMs prototype, engineers shape product strategy—but deep expertise in each discipline is still vital.

10. The Question "How Much Code Is AI‑Generated?" Is Obsolete

The focus shifts to whether code is supervised or unsupervised and whether models can not only add but also delete, refactor, and control complexity. Models tend to add complexity without performing the necessary clean‑up.

11. The Most Valuable People Combine Initiative with Stable Taste

These individuals can discover gaps, make judgments in ambiguity, drive exploration to delivery, and continuously correct direction, regardless of their formal role.

12. Future Products as Work Bases, Not Bigger Apps

Products like Codex will evolve into a work entry point and base that orchestrates other tools (browsers, Excel, Notion, Slack, etc.) rather than trying to replace every specialized application.

13. Practical Advice for Today’s Practitioners

Focus on the unique outcomes you can deliver, not on clinging to existing processes. Adapt workflows, experiment with new methods, and prioritize knowing what’s worth doing, discerning quality, driving things to completion, and finding effective work patterns as tools evolve.