learn-ai-pm-with-phoebe / PM session 1 of 10
Learn AI Product Management with Phoebe · PM track · Session 1 of 10

The job, and a governed setup

AI is very good at the parts of product management that look like writing and very bad at the part that is actually the job. It will draft a PRD from one sentence, and the PRD will read beautifully and decide nothing. This session sets the line: what you delegate, what you never delegate, and what you must never paste into a chat window. Then you meet the product you will make decisions about for the next nine sessions.

🟢 PM track PMs · founders · product leads Bring a real backlog item Start here
0-3 · Welcome 3-18 · What AI does to the PM job 18-42 · The governed setup + meet Cadence 42-45 · Q&A
Part 0

How this track works

Ten sessions, one decision. You are the PM on Cadence, the AI note-taker from the sibling courses - self-serve plus team and enterprise motions, already branded, marketed and financed elsewhere in this series. Nobody has decided what it should build next. That is your job, and this track walks the front half of it: framing an evidenced problem (b2), synthesising research without inventing it (b3), jobs to be done and the opportunity tree (b4), the spec and what makes one signable (b5), prioritisation and trade-offs (b6), metrics and instrumentation before the build (b7), experiments and honest readouts (b8), the roadmap narrative (b9), and one full decision end to end in the capstone (b10).

This course stops where delivery starts. The moment the decision is made, charters, milestones, critical path, risk registers and status reports belong to learn-ai-project-management. Product management decides what to build and why; project management gets it built. b10 hands over explicitly.

Live - presented in session Self-study - read after class ▶ Live lab - scores real text Sources covered
★ What you walk out with today A written line between what you delegate and what you own, a governed setup you can defend to security and legal, and the Cadence brief that every later session builds on.
Part 1 · covers the PM craft literature, AI-assisted knowledge work

Fluency is not judgement 7 min live

The PM job splits into two halves that AI treats completely differently. One half is production - drafting, summarising, restructuring, reformatting, first passes. AI is genuinely excellent at it and you should hand nearly all of it over. The other half is commitment - choosing, cutting, saying no, being accountable when it was the wrong call. AI cannot do any of that, and the failure mode is that its output looks like it did.

Production · delegate nearly all of it first-draft PRDs, specs and one-pagers summarising 30 interviews into themes to check rewriting the same update for three audiences turning a decision into a rollout note finding the counter-argument to your own case Test: could a competent stranger check it in 5 min? Commitment · never delegate any of it choosing which problem is worth a quarter deciding what you are NOT doing setting the number you will be judged on telling a stakeholder no, with a reason owning it publicly when it did not work Test: whose name is on it if it goes wrong? The trap sits between the two columns, not inside either one. Ask for a spec and you get a document with the SHAPE of a decision and none of the substance - a problem with no evidence, a metric with no baseline, no non-goals, no trade-off. Session b5 scores exactly that.
🔍 Click to zoom - two halves, two completely different answers
LiveThe four things AI will invent if you let it4 min

Every one of these shows up in a drafted spec, and every one is a decision that has silently been made for you by a tool that has no stake in the outcome:

  • Evidence. Ask why users churn and you get plausible reasons in confident prose. None of them came from your users. Session b3 makes this failure happen deliberately with real interview snippets.
  • A metric. "Improve engagement" is what a drafter writes when nobody told it what success means. It is not a metric; it is the absence of one, formatted to look like its presence.
  • Scope. With no non-goals in the input, everything adjacent stays arguably in scope, because a drafter has no reason to exclude anything.
  • Consensus. The most dangerous one. AI writes as though the trade-off does not exist, so the document reads like an agreement nobody actually made. You find out in the review, or worse, in the retro.
The one-line rule for this whole course Use AI to produce anything you are willing to check, and never to decide anything you would have to defend. If you cannot check it in five minutes, you are not delegating - you are guessing with extra steps.
Self-studyWhat actually got faster, and what did not3 min read

Be precise about the gain, because the vague version of this claim is how teams end up disappointed:

  • Genuinely faster: the blank page. First drafts, restructuring, summarising, adapting one message for different audiences, and finding the argument against your own proposal. Hours per week, reliably.
  • Not faster, and never was: getting five stakeholders to agree, knowing which of two good options is right for this company this quarter, and the conversation where you tell engineering the date moved. These are the job.
  • Actively slower if you are careless: anything where you have to verify the output more carefully than you would have written it. Research synthesis is the classic - a fabricated theme costs more to catch than the synthesis cost to do.

The honest summary: AI removes most of the typing and none of the deciding. If your week is mostly typing, it will transform your week. If your week is mostly deciding, it will make your documents better and your calendar identical.

Part 2 · the setup you can defend

What you must never paste 6 min live

PMs handle exactly the material that should not go into a consumer chat window: customer interview recordings, named account details, unreleased roadmaps, pricing under discussion, and anything a sales team said under NDA. The setup below is the minimum that survives a conversation with security, and none of it requires a big tooling purchase.

MaterialRuleWhy this line and not another
Customer interview transcriptsde-identify first: no names, employers, or contact detailsyou rarely need identity to find a theme, and consent for research almost never covers a third-party tool
Named account or deal detailnever - describe the segment instead"a 40-seat agency on the team plan" carries the product signal; the account name carries only risk
Unreleased roadmap and datesonly in a workspace your company controlsthis is the leak that shows up in a competitor's deck, and it is your name on the doc
Pricing under discussionnever, in any toolcommercially sensitive and often contractually restricted; the upside of drafting it is near zero
Your own drafts and frameworksfine, and this is most of the valuethe highest-return use is on material you authored, so start there
LiveThree things to fix in your setup this week3 min
  • One place, agreed with security. Not four tools and a personal account. What matters is that the retention and training settings are known and written down somewhere, not which vendor you picked.
  • A de-identification step in the research flow. Do it once, at intake, not per prompt - a habit you have to remember is a habit that fails in the week you are busy.
  • A visible marker for AI-drafted docs. One line at the top: drafted with AI, reviewed by you, on this date. Not ceremony - it tells a reviewer how hard to read, and it means the accountability question never becomes ambiguous.
The disclosure question, answered You do not owe anyone a footnote for a drafted status update. You do owe it when the artifact is evidence for a decision - a research synthesis, a competitive claim, a metric interpretation. In those cases, say where it came from and what you verified, because someone will act on it as if you checked.
Part 3 · the running project

Meet Cadence, and the decision on your desk 5 min live

Cadence records meetings and transcribes them. It sells self-serve to individuals and to teams of 5 to 50 seats, with an enterprise motion above that. It has a brand, a launch campaign and a finance model - built in the sibling courses. What it does not have is a decision about what to build next quarter, and three people have already told you what they think it should be.

Three asks, all reasonable, all from someone senior Sales wants live notes "every demo asks for it" Evidence offered: 3 lost deals Cost: a streaming pipeline we do not have Support wants summaries 214 tickets in 3 months: "cannot find what was decided" Second most common tag Cost: unknown until b6 The CEO wants an agent "everyone is shipping agents" Evidence offered: none Cost: a quarter, at least Hardest to say no to Your job is not to pick the loudest. It is to make the asks comparable. By b6 all three carry an evidenced problem, a named segment, a measurable outcome and a cost - and only then is a decision possible. Two of the three will not survive that process, which is the point of the process.
🔍 Click to zoom - the real starting state of most product decisions
LiveWhat you already know, and what you do not4 min

The signals available on day one, which is fewer than you would like:

  • Support tags: 214 tickets in three months tagged "cannot find what was decided" - the second most common tag.
  • Eight interviews with team admins on paid workspaces, run last month for a different reason. Six of the eight said they never reopen a transcript once the meeting ends.
  • Product analytics: the week-1 transcript revisit rate is 11% of recorded meetings. Nobody has ever set a target for it.
  • A survey: 68% of surveyed admins say they still take their own notes during meetings - in a product whose entire promise is that they would not have to.
  • What you do not have: any idea what a summary would cost to run, whether the 214 tickets came from the same 12 accounts, or whether live notes actually lost those three deals.

That last line matters more than the first four. Most of the work in b2 and b3 is turning "we have some signals" into "we have a problem statement somebody could disagree with on the evidence".

Self-studyBring your own decision2 min read

Every session ends with the same instruction applied to your own work, so pick your case now and keep it for all ten sessions. The best candidate is a backlog item that has been argued about more than twice and decided zero times - the ones where the argument keeps restarting are exactly the ones missing an evidenced problem or an explicit trade-off.

Write down, in one line each: who asked for it, what evidence they offered, and what you would have to believe for it to be the right call. Keep that page. In b10 you will run your own item through the same five checkpoints as Cadence, and the difference between the b1 version and the b10 version is the course.

Try it now

The thing that made this course exist ★ 6 min · everyone

Session b5 is a lab that scores a spec against the six things that predict whether a build goes well. Here is the shortest possible version of why it exists: paste any spec you have written into the box, press Score, and read which of the six it is missing. The scoring runs on the real text, in your browser - nothing is sent anywhere.

Real world

Rung 1 is not a strawman. That first spec - 15 out of 100 - is what a good drafting tool genuinely produces from one sentence, and it is better formatted than most specs that ship. It has a problem section, a solution section, a success section and requirements. It reads like a decision was made. Nothing in it can be checked, nothing is excluded, nothing is measurable, and no alternative was rejected. Teams ship from documents like that every week, and then argue in the retro about what was actually agreed.

Homework

Try it yourself - this week ◐ 20-30 min total

Source material

Sources covered

Full source map in materials/official-course-map.md. This page covers:

The PM craft literature on the role - decision ownership, saying no, accountabilityPart 1 · the production / commitment split
AI-assisted knowledge work - where drafting help is real and where verification costs exceed itPart 1 self-study · the honest gain, stated narrowly
Research ethics and data handling for user researchPart 2 · the de-identification rule; consent frameworks are not this course
Delivery: charters, milestones, critical path, status reportingOut of scope by design - learn-ai-project-management
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Which of these should a PM delegate to AI without hesitation?

Production work - drafting, summarising, restructuring, arguing the other side - is what AI is genuinely good at, and the test is whether a competent stranger could check it in five minutes. Choosing and committing stay with the person whose name is on the outcome.

2 · A drafted spec says success means "improve engagement and increase satisfaction". What has actually happened?

A drafter has to write something in the success section. With no metric definitions in the input it produces a directional phrase - which is why rung 3 of the lab, where the metric definitions are added, is the rung that fixes it.

3 · Sales wants live notes, support has 214 tickets about finding decisions, and the CEO wants an agent. What is the PM's first move?

Two of the three will not survive being made comparable, and that is the process working rather than the process failing. Sessions b2 to b6 are that work; the decision at the end is almost easy once the asks are in the same units.

PM session 1 cheat sheet · pin this

The splitProduction (drafting, summarising, restructuring) delegate. Commitment (choosing, cutting, owning) never.
The one-line ruleProduce anything you will check; decide nothing you would have to defend.
Two testsCould a stranger check it in 5 minutes? Whose name is on it if it goes wrong?
What AI inventsEvidence, a metric, scope, and consensus. The last one is the dangerous one.
Never pasteNamed accounts, unreleased dates outside a controlled workspace, pricing under discussion.
Setup, 3 itemsOne agreed place · de-identify at intake · a visible marker on AI-drafted docs.
Cadence day one214 tickets, 8 interviews (6 never reopen), 11% revisit rate, 68% still take notes.
Where this stopsAt the decision. Delivery is learn-ai-project-management. Next: b2, the evidenced problem.