AI for Project Managers
Most of you knew me when I was a Product Owner. A year ago I would have laughed at the idea of being where I am now. Then I started using AI seriously.
I know some of you have tried Copilot, gotten something garbage back, and concluded the hype is hype. I get it. I'm not going to tell you AI is magic — I'm going to tell you it's a coworker who needs onboarding.
Two minutes. All ten obvious in hindsight.
A real-shaped BRD. Already through one round of internal review.
She's a coworker, not a prompt. She remembers what she's looking for. Every time.
BRDs, PRDs, user stories — they aren't going obsolete. They're the vocabulary AI was built to listen to.
Station One — live visibility for stations — has been in pilot for 2 months. The thing that closed the gap wasn't learning to code faster. It was learning to spec better.
The answer you need exists. It's just scattered across apps that don't talk.
Cowork ships with three drawers. The connectors I need are in one. The skills I'll write are in another. Someone already built the third.
This is the part Copilot can't do — Copilot reads one app at a time.
Memory is what Cowork remembers about you — your role, projects, how you work. Skills are what Claude knows how to do. And unlike memory, skills travel.
A skill is a folder of instructions, templates, and rules. Claude only loads what's needed, when it's needed.
Skills → what Claude knows. MCPs → what Claude can reach. Different things.
Sometimes the missing input is a perspective, not a piece of data.
Save these as personas inside a skill. Now they're always available — without a meeting.
Even with memory, a single thread degrades. Long brainstorms drift. Start fresh threads more often than feels natural — and let memory carry what matters between them.
Details aren't the problem — demands are. Same context, different door: invite questions instead of asking for output.
"Driver app v2 — March 15, team Marie/Sarah/Mike, address-issue handling, risks A through G, deps with sortation + dispatch, success = adoption + CSAT, plus training, rollback, peak prep — please write the PRD."
↳ Door closed. PRD comes back generic. You rewrite for an hour.
"Driver app v2 — March 15, team Marie/Sarah/Mike, address-issue handling, risks A through G, deps with sortation + dispatch, success = adoption + CSAT, plus training, rollback, peak prep. Before we draft — what's missing?"
Claude → Two things: rollback per-feature or all-or-nothing? And — has Marie weighed in on peak-day exception flow yet?
↳ Same context. Door open. The gaps surface before the draft.
Sonnet handles 95% of PM work. Faster iteration = faster learning. Opus is the senior partner you call in for deep reasoning — not your daily driver.
Vague-but-polite asks get vague-but-polite work back. Specificity is respect — for Claude, and for the coworker you'd give the same task to.
"hey claude, could you maybe help me write something about the driver app rollout? thanks!"
↳ You get something. You rewrite it. You wonder if AI is overrated.
"Write a 1-pager for ops leadership on the driver app v2 rollout — what changes Friday, what doesn't, what to tell drivers. Plain language. Keep it warm."
↳ Ships in one round. Politeness costs nothing. Specificity buys everything.
Every project has rules you don't want to re-explain. Put them in the project instructions — Cowork calls it Folder instructions, Claude Code calls it CLAUDE.md, same thing. Then keep it ruthlessly short.
Folder instructions
Use this to give Claude instructions for working in this folder.
Belongs
Doesn't belong
Rule of thumb: under 20 lines. Trim quarterly. Long is the same as no rules.
Pick one. Do it once. That's the whole assignment.
Two surfaces for the part that comes after the talk.