The obituary keeps getting written. "PRDs are dead. Agents will figure it out. Just prompt the coding tool." A few weeks ago Anthropic mentioned they'd built a recent product with zero traditional PRDs, and the takes flowed.
Meanwhile I've interviewed over 70 PMs across companies of every shape this year. The #1 request I got — by a large margin — was help writing better PRDs. Not "should we still write them." Just: help me do this well.
Both things can be true. The reason they're both true is where this gets interesting.
The old PRD had a real problem
Static. Long. Imprecise. You write a document, someone reads a version of it, the requirements shift, the doc doesn't. By the time it hits engineering, half of it is fossil. You either rewrite the plan in Slack or you build the wrong thing. Neither outcome is great.
That was survivable when engineering cycles were measured in weeks. Precision was a luxury; the gaps got closed in the daily standup, or over a whiteboard, or in a 30-minute sync where the tech lead squinted at the doc and said "so what you actually mean is…" You could get away with a lot.
Then the reader changed
When the "reader" of your PRD is an AI agent scaffolding code overnight, imprecision isn't a rounding error. It's the whole delta between "you shipped what you meant" and "you shipped what your document literally said, which is different." Agents don't stop to ask. They don't do the squinting-tech-lead move. They confidently build the wrong thing.
So the PRD didn't die. It got a new audience.
Your PRD now has to work for three readers at once:
- The PM (you) — thinking through what to build, what to cut, what it's competing with in the customer's head.
- The engineer — understanding scope, constraints, non-goals, and the edges where reasonable people would disagree.
- The AI agent — turning the doc into work with no context beyond what's written down.
Each of them needs something slightly different. The PM needs a thinking scaffold. The engineer needs decision boundaries. The agent needs unambiguous scope and explicit non-goals. Miss any of the three and something breaks — usually silently.
Which is why "just prompt it" doesn't scale
You can vibe-code a feature. You can't vibe-plan a roadmap. The moment you're coordinating more than one system, more than one team, or more than one week, someone has to write down what the thing is and isn't. Not because tradition demands it — because otherwise every reader (human or agent) fills the gap with their own assumption. And you don't want that.
What the modern PRD does differently
Four shifts, mostly about what the doc refuses to do:
- Gathers current context first. No writing before knowing what already exists — codebase, prior decisions, customer data. Otherwise your PRD is fiction that happens to be well-formatted.
- Names gaps out loud. Old PRDs tried to look complete. Modern ones explicitly flag: "we don't know how heavy users will use this," "pricing model TBD," "unclear whether this crosses into GDPR scope." An agent that sees the gap can ask. An agent that inherits a papered-over gap builds on top of it.
- Writes for three audiences on the same page. Not three docs. One doc, with a structure that lets each reader take what they need — narrative for the PM, spec for the engineer, unambiguous scope for the agent.
- Section-by-section, not all-at-once. You can't hold a whole PRD in your head. You can hold one section. Modern PRD writing is iterative — thinking mode, then drafting mode, then refining — one part at a time.
About Anthropic
Anthropic didn't ship "without PRDs" the way the post implied. What they did was compress. Fewer artifacts, faster feedback loops, tighter teams. That works when the team is eight senior people who share context effortlessly and can rebuild the shared model in a 20-minute conversation.
If your team is 15+ people, distributed across time zones, coordinating with sales and support and design — you still need the artifact. You just need a better one.
The Smart PRD Skill
I built one. It's a Claude skill that walks you through writing a modern PRD the way I'd run the session with you: gathers context first, asks section-by-section, writes for all three readers, and flags gaps instead of papering over them.
You install it once. It runs in every PRD session after that.
Grab the skill → No upsell. Just the download.