The difference between a PM who ships and one who drowns in busywork is often not intelligence or work ethic. It is whether they have learned to use AI for the repetitive, high-friction parts of the job. I have tested hundreds of prompts across customer research, data analysis, strategy, writing, and prototyping. The 20 in the source post are the ones that actually save hours and produce outputs good enough to ship without rework. Three of them save the most hours.
The three highest-leverage prompts
The JTBD extractor (Prompt 1) analyzes a customer interview transcript and pulls out every distinct job to be done, functional, emotional, and social, with the exact quote that reveals it, a job statement, and a ranking by importance, output as a markdown table. It is the foundation of good positioning and roadmap work. The pro move is pasting three or four transcripts from the same segment at once so the AI flags the consensus jobs, which matter more than outliers.
The ticket-theme clusterer (Prompt 2) takes a CSV of support tickets and identifies the top five to seven themes ranked by frequency, revenue impact, effort, and impact if shipped. It turns thousands of scattered tickets into a prioritized roadmap input and saves hours of manual clustering. Include customer MRR or ARR in the CSV and the AI automatically weights high-value requests higher, which is what you should be doing anyway.
The one-pager-from-notes generator (Prompt 14 in my ranking) drafts a full spec from rough notes so you stop writing 40 pages. Between them, these three replace the bulk of the weekly grind that eats a PM's calendar.
Why they work
Three patterns lift a prompt from generic to surgical. First, put the role, audience, and constraint in the first sentence, like "you are a senior PM at a B2B SaaS preparing a 200-word board update." Second, paste real artifacts, actual transcripts and ticket exports and competitor pages, instead of describing them. Third, demand a specific output structure so the result is drop-in usable. Every strong prompt in the library follows this.
The line you do not cross
A good PM with a good prompt library does in two hours what a PM without one does in two days. But the job becomes higher-leverage, not obsolete. Never automate the judgment or the relationships. Do not ship decisions you will own without reading every word, and do not let AI message customers or executives without your review. The rule of thumb: if a human will be hurt or misled by a bad output, you have to be in the loop.
Pick one thing to try this week: take your last customer interview transcript, run the JTBD extractor on it, and compare the structured output to the notes you took by hand.