AI for product managers: PRDs, research, roadmaps

Most PMs reach for AI to write PRDs first, which is exactly the wrong place to start. A PRD is 10% writing and 90% having the right thinking. AI is great at the 10%. The 90%, the part where you decide what's actually true about the user and what trade-off you're willing to defend in a room of engineers, that's still your job.

The good news: there are three other places where AI saves a PM real hours, and most PMs ignore all three.

1. Synthesizing user research

This is the biggest unlock and almost nobody runs it well. You did ten interviews. You have ten transcripts (or you used Granola, or your researcher exported from Dovetail). The default move is to skim them on a Sunday and convince yourself you spotted patterns. You didn't. You spotted the two quotes you remember.

Paste all ten into Claude. Long context handles it. Ask for the three most-mentioned pains, with three verbatim quotes per pain and the speaker's role attached. Then ask for the contradictions, the places where two users said the opposite thing. That second ask is where the value lives.

The few-shot pattern from the marketer guide applies cleanly here. If you've written a synthesis you were proud of (the kind your design partner forwarded around), paste it as a sample and ask for the same shape on the new transcripts. Two examples of "good output" beats a paragraph describing what good looks like.

NotebookLM is a decent alternative if you want a chat interface across a corpus you'll keep adding to. Claude wins when you want one-shot synthesis with structure.

2. PRD drafting from rough ideas

Once the thinking is done (problem clear, user clear, constraints clear), the document itself is mostly mechanical. ChatPRD is purpose-built for this and worth the subscription if you write more than two PRDs a month. If you don't want another tab, paste two of your past PRDs into Claude as voice examples, then drop in your bullet-point thinking and ask for a draft in the same shape. Notion AI is fine if your specs already live there and you don't want to copy-paste, just don't expect it to push back on weak reasoning.

A working prompt:

Below are two PRDs I wrote that shipped successfully [paste both]. 
Match the structure, voice, and level of detail.

Here's the rough thinking for the new feature:
- Problem: [3-5 bullets]
- User: [who, with one quote from research]
- What we're considering building: [bullets]
- Open questions I want engineering to weigh in on: [bullets]
- Out of scope for v1: [bullets]

Draft the PRD. Where my thinking is thin, mark it [TODO: weak] 
instead of papering over it.

That last instruction is the one most PMs skip and the one that matters most. You want the model to flag where you haven't thought hard enough, not to paint over it with confident prose.

3. Roadmap narratives

The quarterly roadmap doc, the all-hands slide, the email to the exec team. The decision about what's on the roadmap is yours. The story you tell about why the items hang together is mostly assembly work.

Paste your Linear or Jira items, your themes for the quarter, and one or two prior roadmap write-ups in your voice. Ask for a coherent narrative grouped by theme, with one sentence per item connecting it to the theme. This is for the write-up, not the prioritization decision. If you let AI do the prioritization, you'll get the average opinion of the internet about what good products look like, which is not a roadmap, it's a beige.

The PM-specific trap

AI hallucinates plausibly about your product. Ask "what should we build next?" cold and you'll get a confident answer that has nothing to do with your actual users, your actual data, or the three constraints your CTO will surface in the first five minutes of the review.

The fix is the same in every prompt: anchor in real artifacts. Transcripts. Tickets. Usage data exports. Support themes from the last 30 days. The model is a synthesis engine, not a strategy engine. Feed it your reality and it's useful. Ask it to invent your reality and it will, badly.

What AI is bad at, for PMs

Prioritization. It will average opinions and hand you a stack rank that offends nobody and excites nobody. The hard call between two equally-loved features, the one where two of your best engineers disagree, that's not a tokens problem, that's a judgment problem. Anything political: which VP needs to be in the loop early, why marketing is quietly upset, who actually owns the decision. The model does not know your org and pretending it does is how PMs get into trouble.

Keep those for yourself. Hand the model the transcripts, the drafts, and the roadmap write-up. Save the four hours.

Up next

The next guide swaps the spec for the canvas. Pillar 2 guide 5, "AI for designers: from blank canvas to brand-consistent assets," covers how design teams use the same stack to ship work that doesn't look like every other startup's Midjourney moodboard.