Andrew Ambrosino, who leads product and engineering for the Codex app at OpenAI, says implementation is no longer the expensive part of building, so taste is now the scarce skill.
Ambrosino argues PRDs are not dead, saying the right medium depends on the point being made, and warns that polished prototypes can be mistaken for production-ready work.
Ambrosino says nine-month plans are false precision, citing the Codex app, which he says would have failed had it shipped in November because the models were different.
When anyone can build, taste becomes the scarce product skill
AI has made building cheap, which flips product work from careful de-risking before building to building first and curating what comes out. That makes taste the scarce skill and makes long plans unreliable, because model capability shifts what can work from one month to the next. Codex is OpenAI's bet that one app can become the home base for this way of working.
Takeaways
Implementation is cheap, so taste is scarce. Anyone can build a feature by talking to models, and OpenAI staff get unlimited tokens. Roughly 90 teams may build the same idea at once. Taste means curating many attempts, with aesthetic, systems-thinking and presentation parts.
The product process has inverted. Teams used to de-risk implementation through documents, research and prototypes. That assumed building was expensive. Now people start from implementation. Ambrosino agrees with Claude Code's design head Jenny that the design process is dead, though it was always somewhat academic.
PRDs are not dead, but medium must fit the point. A document suits product clarity around a vague area, while a prototype suits testing an interaction. A polished prototype can look production-ready, so teams must say what stage the work is at.
Long plans fail because models keep changing. Nine-month plans are false precision. The Codex app released in February would have failed if shipped in November, with only the models different. The early web Codex was too early, while Claude Code worked because models were ready.
Roles are collapsing but disciplines still matter. The Codex team has designers who speak engineering and product managers who write code. Ambrosino opposes abolishing product as a role, since real disciplines would be lost, and says managers remain because not everyone can work on everything.
Models lag at design for practical reasons. Design is harder to grade than code, human taste sits in the feedback loop, and labs invest in what speeds AI research. Ambrosino expects improvement, but the link between software design and code, and novelty, remain harder.
Codex is becoming a general home base. The in-app browser started as a developer testing tool and was rebuilt on the Atlas engine. The vision is a place to start, track and automate work, connecting to tools like Excel and Premiere Pro, with plans to add Codex to ChatGPT and Atlas.
How Codex is used inside and outside OpenAI
Nearly all OpenAI employees use Codex weekly, including non-engineers in marketing, finance and legal.
Under last year's standard, 100 percent of the product is AI-written. The live question is supervised versus unsupervised writing, and models still tend to add complexity.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Ambrosino runs a daily brief drawn from more than 3,000 Slack channels. He wants personal automations like this to become product features.
A videographer and Codex built a Premiere Pro extension, though Codex only edits files behind the tool and is not a video editor.