Meta Replaced the PRD With a Morning Dashboard. Here's What That Costs You.

"Six months in, it's accurate. At the start, it wasn't." That's Jagjit Chawla, the Meta executive who leads Feed, Reels, and Search for Facebook, describing the system he built to replace his own morning status report. It runs overnight against his engineering partners' code diffs, emails, and chats, and produces a dashboard by the time he wakes up: decisions the team made, decisions waiting on him, and outstanding asks from people above or beside him he hasn't answered yet. He didn't buy this – he built it, then spent half a year fixing it before he'd act on what it told him.

That six-month detail is the part I keep coming back to, because most of the AI-and-PM coverage right now skips straight past it to the flashier claim: the PRD is dead. Chawla's version of that claim, on Nikhyl Singhal's "The Skip" podcast, is specific and worth taking at face value – but it's an account of his own team, not a Meta-wide mandate, and I haven't found evidence any other big tech company runs something equivalent. Treat it as one credible practitioner's account of where this is heading, not as proof everyone's already there.

What actually replaced the document

Two years ago, in Chawla's account, a PM's output was a real PRD – problem statement, feature set, hypothesis, the whole document. Today, on his team, that's compressed to one paragraph describing the problem, paired with a working prototype, and for anything ML-heavy, an eval set. The job moved with it: instead of writing the document alone at a desk, the PM sits next to the engineer during the eval and says yes or no – is this a good outcome, or a bad one. Chawla calls that "what real product management looks like right now."

Notice what didn't disappear in that description. Judgment didn't go away; writing did. The PM used to demonstrate judgment by producing a thorough document that proved they'd thought it through. Now they demonstrate it in real time, standing next to working code, making a call. That's a faster loop, and it's also a much less forgiving one – there's no draft to revise quietly before anyone sees your reasoning. You're judging out loud, immediately, or you're not doing the job.

The dashboard is the same trade, one level up

The morning-report system is the executive version of the same swap. Before, information about what the team was doing traveled up through management in what Chawla calls a "compression algorithm" – each layer summarizing for the layer above, arriving at the top as a thin, filtered sliver. His agentic dashboard replaces that chain directly: it reads the raw material – diffs, chats, documents – and surfaces decisions made, decisions pending, and VIP asks nobody's answered, without a human doing the summarizing in between.

That's a real efficiency gain, and it's also why the six months matters so much. A system reading raw diffs and chats has to learn, specifically for this team, what counts as a decision worth flagging, who actually qualifies as a VIP whose ask can't wait, and what "accurate" even means in a stream of overnight engineering activity. None of that is generic; it's the exact tuning work a human compression chain used to do implicitly, through years of institutional knowledge about who matters and what's urgent. Chawla just did that tuning explicitly, over six months, instead of inheriting it slowly the way a new manager does.

The same trade shows up in how Meta ships, not just how it reports

Chawla gives one more example that I think is the clearest illustration of the whole pattern, because it's the least abstract. At roughly two billion daily users, Meta gets tens of thousands of bug reports a day – more than any human team could read, so for years, effectively, nobody fully did. Now a chain of AI agents categorizes each report, validates it against screen recordings, proposes a fix, and routes it to the right engineer for review before it ships. Work that was previously unworkable at that scale becomes routine.

But Chawla is upfront that ramping AI-assisted code generation also increased site incidents – SEVs, in Meta's internal shorthand – and he's specific about why the stakes are unusually high there: at Meta's scale, even a 0.1% crash rate reaches tens of millions of people. Meta's response wasn't to slow down; it was to build additional AI-based validation layers to catch what the faster pipeline introduced – AI creating a problem and AI helping solve it, in the same breath. Nobody in this account claims the incident rate is now fully under control, and the source gives no clean before-and-after number. What it does show is the same shape as the PRD and the dashboard: a process gets dramatically faster, and the judgment that used to be exercised by sheer human bandwidth has to be rebuilt somewhere else, deliberately, or the speed just ships more mistakes faster. It's a live version of the same displacement I've written about in whether product managers are actually becoming obsolete – the role isn't vanishing, the judgment it used to hold is migrating into new checkpoints, and someone has to build those checkpoints on purpose.

Where I think most teams will get this wrong

The mistake I'd bet on is skipping straight to the dashboard and assuming the six months was optional overhead rather than the actual product. A team that stands up an equivalent system in a week and starts trusting its output immediately hasn't built what Chawla built. They've built something that looks the same and hasn't been taught yet what a real decision looks like on their team specifically – which means it's making that judgment call with none of the tuning, at the exact moment leaders are inclined to trust it most, because it's new and it's a dashboard and dashboards look objective.

This connects to something I keep circling back to across a lot of this AI-adoption coverage: the judgment work doesn't get eliminated by systems like this. It relocates. In Chawla's case, it moved from "write a thorough document" to "sit next to the eval and call it live," and separately, from "climb the management chain slowly" to "spend six months personally teaching a system what matters." Both moves are real productivity gains. Neither one is free, and the org that treats either as a checkbox – ship the eval process, stand up the dashboard – without doing the underlying judgment-transfer work is going to get a fast, confident, wrong answer instead of a slow, hedged, right one. I've written elsewhere about what happens to a product strategy once the thing you're building starts making its own calls – this is the same problem showing up one layer earlier, in how the organization decides things, not just in what the product does.

What this is actually evidence of

I want to be careful not to oversell this as a trend, because it isn't one yet, at least not one I can point to outside this single account. What it is: a specific, credible, on-the-record description of what it looks like when an organization takes the "AI compresses PM work" claim seriously enough to rebuild the actual reporting mechanism around it, rather than just using AI to draft PRDs faster. That's a much bigger commitment than most of the "AI for PMs" content out there is describing, and it's why I think it's worth more attention than the headline version of this story – "Meta's PRDs are one paragraph now" – usually gets.

The PRD didn't get shorter because writing got automated. It got shorter because someone decided the judgment belonged somewhere more immediate than a document, and then did the unglamorous work of making that true. Most teams reading this story will copy the paragraph. Very few will do the six months.