How to write a PRD in the AI era

The PRD is the most-written document in product management. Every quarter, every team, every PM writes them. The shape of one (context, problem, hypothesis, scope, success metrics, risks, rollout) has been roughly the same since the early 2010s.

AI hasn't actually changed how most people write them.

What AI has changed is the floor. ChatGPT has made it possible to produce a PRD-shaped document in five minutes that contains every section, reads grammatically correctly, and is essentially worthless. The form is right. Everything else is generic. The personas are generic. The risks are generic. The success metrics are generic. The scope is generic. It could be any company shipping anything.

The reason is simple. Generic AI doesn't know your company. Every prompt starts from zero. You ask write me a PRD for a referral program and you get a referral program PRD that any company in the world could have written.

That's not a PRD. That's a PRD-shaped artifact.

The 80/20 of a real PRD

A real PRD is 80% context and 20% writing.

The 20% is phrasing the problem, structuring the doc, drafting the success metrics. It's what most people focus on when they think about writing. It's also what generic AI is decent at. Form, prose, structure.

The 80% is knowing what the customer actually said. Knowing what your team has already tried. Knowing what the data shows. Knowing why you decided against this same idea two quarters ago. It's what makes the doc useful. And it's exactly what generic AI does not have.

When you skip the 80%, your PRD reads like a template. When you bring it in, your PRD reads like a decision.

The leverage isn't faster writing. It's faster grounding.

What “context” actually means

When PMs talk about contextthey usually mean a Slack thread or a stakeholder email. That's information, not context. Context is the structured set of things your AI tool can actually pull from when it drafts.

For a PRD, the relevant context lives in four places.

The customer voice. Recent interview transcripts. Support tickets that mention this surface. NPS verbatims. Anything that tells you what your real users are saying about this problem in their own words.

The personas.Not the slide-deck personas from the seed round. The current ICP profiles, with their actual jobs to be done, the channels they use, the words they use, what they pay for, what they don't.

The historical decisions.Why did the team decide against this same idea last quarter? What are the prior PRDs in this surface area? What's been promised in the roadmap and not yet delivered?

The data. Conversion funnels, retention curves, segment breakdowns. The numbers that any honest scoping discussion has to start from.

These are the inputs that make a PRD worth writing. If your AI tool doesn't have access to them, no amount of clever prompting saves the output.

The wrong workflow

Here's what most PMs are doing in 2026.

They open ChatGPT or Claude. They paste the rough shape of an idea. They write a long prompt that recapitulates what their company does, what their users want, what the goal is, what the constraints are. They get back a draft. They edit it. They paste it into Notion or Confluence.

The PM has spent thirty minutes typing context that the AI immediately forgot. Tomorrow they'll do it again. The week after, they'll do it again. The context isn't accumulating anywhere. Every PRD starts from zero.

That's the workflow that produces PRD-shaped artifacts.

The right workflow

You start with the question, not the prompt. Should we build a referral program for the Pro tier?

You ground that question in the four sources above. Pull the recent customer interviews mentioning friction upgrading from Free to Pro. Reuse the existing ICP personas. Surface prior decisions about acquisition versus retention, and any earlier referral discussions. Pull conversion data from the Free to Pro funnel. Note what direct competitors charge for equivalent programs.

Then you draft. Now the draft is grounded. The personas in the doc are your personas. The risks in the doc are the risks you've actually seen. The success metrics map to your funnel. The scope reflects your engineering capacity.

The output looks like the same PRD shape. The difference is invisible until someone reads it. They'll read it and recognize their own company in the doc. Decisions get made faster. Engineering pushes back less. Stakeholders argue with the substance instead of I think you're missing the customer.

Where Inversion fits

The reason I'm building Inversion is that the right workflow is currently almost impossible. The four sources live in different tools. Customer interviews live in a transcript service. Personas live in Notion, or a deck somewhere, or in someone's head. Historical decisions are scattered across Slack and ten different Google Docs. Data lives in Postgres, behind a SQL query you don't know how to write. Competitive intel lives in your head.

Inversion puts those four sources in one place and lets the AI draft from them. The Brain stores PRDs and decisions and auto-reviews them for staleness. Personas are first-class objects, not slide-deck artifacts. The Data Explorer asks Postgres in plain English. Competitor research is a surface, not a folder. Idea Lab walks an idea from seed to brief through structured gates that pull all of the above.

When you write a PRD in Inversion, you don't paste context. Context is already there. The draft starts grounded.

That's what AI in product management is supposed to feel like. The PRD-shaped artifact problem disappears. The 80% gets done at the speed of the 20%.

The test

If you write two PRDs this month, here's the test. Pick the one you're proudest of. Read the personas section. Read the risks section. Read the success metrics section.

Could any company in the world have written those paragraphs?

If yes, you wrote a PRD-shaped artifact. Generic AI will write the next one for you in five minutes flat. If no, you wrote a PRD that earned the engineering team's time.

The job of a tool like Inversion is to make the second outcome the default, not the exception.

Stop typing your company into a chat window every Monday. The 80% wants to be done.

Nunes