Andrew Ambrosino, product and engineering lead for the Codex app at OpenAI, discusses how AI is fundamentally reshaping product development workflows, with 90% of OpenAI employees using Codex weekly and the app seeing 6x growth since January to over 5 million weekly active users.
The inversion of product development process
Implementation has become cheap while curation and taste have become the expensive, valuable parts of product work
Where traditional product development followed research → ideation → prototyping → implementation, now anyone can build anything immediately
This creates an environment where 90 different uncoordinated teams might implement the same feature simultaneously
The challenge shifts from building to deciding what’s good among many implementations and how to fold insights together
Documents versus prototypes in the AI era
Both documents and prototypes remain valuable but serve different purposes in the product development process
Prototypes work best for stress-testing interaction patterns and getting hands-on feedback
Documents work better for achieving product clarity around vague or complex areas
The key is matching the medium to the specific point you’re trying to make rather than defaulting to one approach
The “primal mark” concept applies: the first artifact created becomes what everyone responds to, so choosing the right starting point matters
What “taste” means in product work
Taste encompasses aesthetic judgment, systems thinking, understanding where the product is going, and how to present ideas effectively
It includes knowing what interaction animations fit the semantic meaning they’re supposed to convey
More importantly, taste involves deciding what to build and how to get there when anything is possible
Good taste determines what’s signal versus noise in a world of infinite content generation
Why AI still struggles with design
Design is harder to grade than software because human taste is part of the feedback mechanism needed for training
Labs historically invested in coding capabilities because they directly accelerated AI research, unlike design skills
Design requires novelty and cultural awareness that differs from software engineering’s preference for established patterns
There’s an abstraction layer challenge: visual design must connect to underlying code architecture and shared components
A rebrand shouldn’t require updating 263 components individually but understanding the semantic relationships between them
The evolving design process
The traditional design process assumed expensive implementation and exhaustive upfront research
Now that implementation is abundant, the process must adapt rather than disappear entirely
Companies creating “baby versions” of products (simplified codebases) allow for rapid vibecoding exploration
The design process lives on but requires clearer communication about what stage an artifact represents
Polished prototypes can mislead stakeholders into thinking features are ready for production when they’re still exploratory
Team structure and zone defense
The Codex team operates with double-digit engineers, roughly half that number on design, and a few product people
Everyone on the team demonstrates agency and taste, with many former founders or people doing founder-shaped work
Teams use “zone defense” for product work: spreading out to cover gaps rather than overlapping closely
Product people aim for company coverage by identifying who’s best at what and creating space between them
Hiring focuses on engineers who are product-minded to maintain product coherence without heavy review processes
IC and management convergence
Individual contributors now manage agents and work rather than typing code character by character
Management happens at different granularities but both IC and management roles involve coordination
The most valuable people can take ideas from conception to completion with good taste and high agency
Command over discipline plus taste to distinguish signal from noise defines success in this environment
Planning in an AI-accelerated world
Planning works best with high-level vision for long-term goals and detailed plans only for short-term execution
Precision in 9-month plans creates false precision since model capabilities shift rapidly
Features must be prototyped to test against future model improvements rather than relying on static planning
The Codex app released in February would have failed in November due to model capability differences
Some features need to be released multiple times as models improve before finding product-market fit
Building for future model capabilities
Teams build features that don’t work yet, treating them as artifacts to test against future model improvements
The in-app browser, computer use, and artifact creation features represent this “build ahead of readiness” approach
Being too “AGI-pilled” early can hurt adoption; matching ambition to current model capabilities matters
Features like the in-app browser required multiple iterations with different intelligence levels to succeed
The latest frontier: autonomous development loops
The question has shifted from “how much code is AI-written” to “is it supervised versus unsupervised”
Current explorations focus on autonomous software development and codebase garbage collection
Models tend to increase complexity rather than reduce it, making true autopilot development challenging
Teaching models which features to build, ignore, or reframe remains an unsolved problem
The abstraction layer between features and codebase organization still requires human judgment
Vision for Codex as a general work platform
Codex evolved from a CLI developer tool to a desktop app aiming to be the best ever created
Internal testing revealed PMF across engineering and research workflows before public release
Non-technical teams (marketing, comms, finance, legal) adopted the app despite it being designed for developers
The vision is a home base for work: starting, ending, and automating tasks across different surfaces
Rather than forcing everything into one interface, Codex connects to existing tools like Excel through connectors
Creative use cases and extensibility
A videographer built a Premiere Pro extension using Codex to automate video editing tasks
Codex understood the user’s tool (Premiere Pro) and created extensions to bridge capability gaps
Two models emerge: seamless interaction with existing tools versus bringing web apps into Codex
Personal workflows vary widely but reveal patterns that could become first-class product experiences
Memory features and mind palace concepts are being explored to reduce individual setup burden
Lessons from failure and career journey
Ambrosino’s startup experience involved years of constant failure in heavily regulated industries
Multiple micro-failures occurred during attempts to merge Codex lessons with ChatGPT
OpenAI’s culture embraces direct feedback through 2,000-message Slack threads criticizing product decisions
Success came after 10-15 years of learning, emphasizing persistence and continuous adaptation
The key advice: don’t get married to exact processes but to outcomes you can uniquely deliver
