At Meta, a Product Spec Is Now One Paragraph and a Prototype

The product requirements document is mostly gone from Jagjit Chawla’s corner of Meta. In its place, a product manager writes one paragraph describing the problem and pairs it with a working prototype. Chawla, a VP of Product who runs a growing share of the Facebook app, including Feed, Reels and Search, walked through the change in a conversation with Nikhyl Singhal on The Skip.

For machine-learning-heavy work, the paragraph and prototype come with an eval set, a battery of test cases run again and again because the models don’t give the same answer twice. “Now you sit next to your engineer on the eval, and you agree or disagree whether this is a good outcome or a bad outcome,” Chawla said. “That is what real product management looks like right now.” Teams work in pods of five people sitting together instead of trading documents.

An agent instead of the org chart

Chawla described the old flow of information as a “compression algorithm.” Engineers know what they checked in, their managers summarize it, and by the time it reaches a VP with fifty projects, each one has been squeezed to a sliver. Now an agent reads every overnight code diff, every waiting email and chat and every document he needs to review, then hands him a red-yellow-green punch list at 7 a.m. In parts of his org it also writes the product review doc, flags the questions worth debating and names who should be in the room.

He built the system himself from generally available tools. Each night it compares the comments he actually left on documents with the ones it predicted and adjusts, and he said it took about six months to get accurate. His advice to everyone else was to pick any tool and start. “Find a scissor. Doesn’t matter what scissor,” he said. The catch is that it only works because Meta writes everything down.

Economists over engineers

Chawla came up at Google believing a PM needed an engineering background. On his team now, the PMs pulling ahead include PhDs in economics and physics who reason from first principles. Where an engineer asks how something will be built, his economists ask what a change does to the incentives of creators, users and advertisers. Fluency with AI coding tools is the floor, he said, not the bar.

Meta made room for the shift by canceling every meeting for a week so the whole org could learn the tools and build with them. That week turned up gaps, such as design tools producing prototypes that broke Meta’s design system, and led to “AI captains” for each function, always picked from inside the company.

The products changed too. Bug reports filed by shaking a phone in the Facebook app, tens of thousands a day, now go to agents that triage them, check the attached screen recordings and draft fixes for an engineer to approve. Not everything went smoothly. As AI-written code ramped up, site incidents rose, and Meta answered with safeguard agents that check changes before they reach production.

Chawla’s own bottleneck has moved as well. His team can produce a fifty-page deck for any idea in half an hour, and he gets five a week. “Now, I am the bottleneck,” he said.