If you read The PM Skill Map last month and walked away thinking "great map, now what do I actually do," this is the answer.
Ten plays. Each one is a workflow I or someone I trust runs every week. Each has a copy-paste prompt skeleton you can adapt to your context, a named anti-pattern, and a verification step. They are not everything a PM can do with AI — they are the highest-leverage ten. Get fluent in these and your hours-per-output ratio shifts in a way that compounds.
One principle first, because it determines whether any of these plays land.
Outsource effort. Keep judgment.
The PMs who get the most out of AI aren't the ones who use it most. They're the ones who know which parts of their job are effort — drafting, formatting, summarising, querying — and which parts are judgment — choosing, framing, deciding, owning. They outsource the first and protect the second.
Drafting documents. Summarising calls. Writing SQL. Extracting themes. Generating variants. Categorising bug reports. Formatting outputs. Reading long things.
Which problem to solve. Which framing to ship. Which tradeoff to make. Which signal matters. Which user to believe. Which risk to take. What "good" looks like.
Every play below is on the effort side of that line. The verification step is what keeps you on the judgment side. Skip the verification and the play stops working in about a week.
The ten plays
The Customer Feedback Synthesizer
You're a senior product analyst. I'll paste [N] [interview transcripts / tickets / NPS comments] below. For each one, extract: - The job the customer was trying to do - The pain or friction they described - The workaround they're using today - One direct quote worth saving Then synthesize across all of them: - Top 3 themes by frequency × severity - Disconfirming evidence for our current roadmap - Two surprising patterns I should investigate Format: a table for the per-item extraction, then a one-page synthesis below. [paste content]
The PRD First Draft
You're an experienced PM at [company type]. I'll describe a problem and a proposed solution. Draft a PRD with: 1. Problem statement (1 paragraph) 2. Goals & non-goals (3 bullets each) 3. User stories (5-8, in INVEST format) 4. Acceptance criteria per story (Given/When/Then) 5. Edge cases (5+) 6. Open questions for eng + design 7. Success metrics (leading + lagging) Be skeptical. Push back where the logic is thin. Problem: [paste] Proposed solution: [paste]
The User Story Writer
Convert the following feature description into 5-10 user stories in INVEST format. Each story:
- Independent
- Negotiable
- Valuable (state the user value explicitly)
- Estimable (rough size tag: S / M / L)
- Small (fits one sprint)
- Testable (acceptance criteria in Given / When / Then)
Include 2 edge-case stories I might miss.
Feature: [paste]
The SQL Translator
I have a [Postgres / Snowflake / BigQuery] table named [name] with columns: [list columns + types] Write a SQL query to [specific question, with metric definition]. Make it readable, comment the joins, include a LIMIT for testing, and use CTEs where it helps clarity.
The Funnel Diagnostic
You're a senior PM. Here are funnel metrics for the last 4 weeks: [paste data table] Step [X] dropped from [Y%] to [Z%]. Generate: 1. Five hypotheses for why, ranked by likelihood 2. For each, the data I'd pull to confirm or rule out 3. For the top two, an experiment I could run in under a week 4. One non-data step worth doing in parallel (user interviews, etc.)
The Competitive Scan
I need a competitive scan of [category] for [target buyer]. For each of these competitors: [list 5-10] Extract: - Pricing model - Their positioning statement, in their own words - Strongest feature - Weakest feature - Any pricing or product change in the last 90 days - Any hiring signal that suggests direction Then synthesize: - Where the market is moving - Which two competitors are on the same trajectory - Who is moving differently Format: table + one-paragraph synthesis per competitor.
The Executive Briefing
You're a strategic communications writer. I need to brief a [CFO / CEO / Board] in 5 minutes on [topic]. Output: 1. One-line top line (what they need to know) 2. Three supporting points, each one sentence 3. The one risk I'm worried about 4. The decision I need from them 5. The full one-page version, if they ask Context: [paste your thinking]
The Bug Triage
Categorize the following bug reports. For each one extract:
(a) Component
(b) Severity (P0 / P1 / P2 / P3)
(c) One-line summary
(d) Is this a duplicate of an existing pattern?
Then group by component and severity. Flag the top 5 to look at first, ranked by: severity × user count × revenue impact.
[paste bug reports]
The A/B Test Designer
I'm running an experiment: Hypothesis: [paste] Change: [paste] Design it for me: 1. Primary metric (one — no more) 2. Guardrail metrics (3) 3. Sample size (baseline conversion [X]%, MDE [Y]%, alpha 0.05, power 0.8) 4. Allocation (50/50 or otherwise, with rationale) 5. Runtime estimate 6. Early-stopping criteria 7. Two plausible ways this could give a false-positive 8. What I'd do if it wins / loses / is inconclusive
The Meeting-to-Action-Items Engine
From these meeting notes, extract:
(a) Decisions made — with the one-line context for each
(b) Action items — with owner and due date
(c) Open questions — that didn't get resolved
(d) Watch-outs — anything I committed to that I might forget
Format:
DECISIONS
ACTIONS (owner · item · due)
OPEN (questions to revisit)
WATCH OUT (my commitments)
[paste notes]
A week in the life
To make this concrete, here is one of my own weeks — lightly edited — with the plays mapped to the actual hours. The point isn't that you should copy the week. It's that once these plays are muscle memory, they slot into a normal PM week without adding hours. They subtract them.
A real PM week, with the plays slotted in
Hours saved are estimates against my pre-AI baseline two years ago. They underestimate; the compounding is real.
Roughly twenty hours back, per week, doing the same job to a higher standard. The compounding shows up in the work that doesn't exist in the calendar — the strategy thinking, the customer calls, the time to actually read.
The stack — what to run all of this on
None of these plays require enterprise software. They require one or two well-chosen tools used well. Here is the minimum viable PM-AI stack as of mid-2026.
| Layer | Tool | What it's actually for |
|---|---|---|
| LLM | Claude or ChatGPT (pick one and go deep) | All of Plays 01-03 and 05-10. Pay for the paid tier. The free model is for tourists. |
| Coding / data | Cursor or Hex with text-to-SQL | Play 04. Pasting SQL between a chatbot and your warehouse is for 2023. Live connection saves the verify step half its time. |
| Meeting capture | Granola, Fellow, or Otter | Feeds Play 10. Transcript-grade notes are now a commodity; pick one and use it consistently. |
| Customer voice | Dovetail or raw transcripts → LLM | Powers Play 01. If you have Dovetail, use it. If you don't, raw transcripts in a folder + a paid LLM gets you 80% of the way. |
| Competitive watch | Crayon / Klue, or DIY with Browserbase + LLM | Powers Play 06. The DIY version is what I run — half a day to set up, then it runs forever. |
| Experimentation | Statsig, Eppo, or LaunchDarkly | Plays in Play 09. The eval scaffolding here matters as much as the model. |
| Knowledge | Notion or Confluence + LLM Q&A | Where your PRDs, RFCs, retros, and customer-research artefacts live. AI gets useful in the company on day 90, when there's enough context for it to be wrong about. |
Anti-patterns to avoid
What separates the PMs getting twenty hours back per week from the PMs getting frustrated is not the tools. It's whether they avoid the predictable ways AI workflows fail.
Using AI to draft what you can't yet think
If you can't sketch the answer on a napkin, AI can't help you. It is an amplifier of clarity, not a replacement for it. The output looks fluent and is therefore dangerous — you ship the model's version of "good enough" instead of yours.
Skipping the verify step
Every play has one. Every play. The plays stop working in about a week if you don't verify, because you build trust in outputs that haven't earned it. Verification is the practice that keeps you honest about what the model can and can't do this month.
Outsourcing judgment by accident
Asking the model "which feature should we prioritize" is not a play. It's an abdication. The model has no skin in your roadmap. Use it for the options; keep the choice.
One-shot prompting
The first response is rarely the best. Push back, ask for variants, ask for the opposite take. The PM who treats prompting as a conversation outperforms the PM who treats it as a vending machine, every time.
Pasting confidential data into a consumer tool
Customer names, revenue figures, internal documents — into the right model with the right enterprise terms, fine. Into a free consumer tier, not fine. Read your tool's data policy once. It will save you a meeting with legal later.
Hoarding the workflow
The plays only compound if your team learns them too. Pair on the prompts. Share the templates. The best PM AI workflows I've seen are team workflows — three or four people getting better at the same plays — not lone-wolf magic tricks.
The closing principle
The PMs I watch get the most out of AI share one habit. They treat their own prompts the way a good engineer treats their own functions — they iterate on them, version them, share them, and retire them when better ones come along. The prompt is the artefact. The output is just what the artefact produced this time.
So the homework, if you've made it this far, is small. Pick one of the ten plays. The one that maps to the thing on your calendar that hurts most this week. Run it. Verify it. Iterate the prompt twice. Save the version that worked into a personal "plays" doc.
By the end of the quarter you'll have ten of these. By the end of the year you'll have built a private operating system that compounds with you, week over week, that nobody can take away. That is the real return on AI for a PM in 2026. Not the demos. The plays.