Sixty product managers at Figma share one library of AI skills, and it quietly rewired how they work. It’s called PM OS, it packs 31 skills, and it started as a side project one PM pitched internally. Here’s how it works and how to build your own version this week, even if you’ve never touched an AI agent.
What PM OS actually is
PM OS isn’t a product you can buy. It’s an internal system built at Figma, and it came from an obvious place: PMs were drowning. Status updates scattered across five tools, context buried in meeting notes, experiments tracked in yet another dashboard. Everyone agreed AI should help, but the tools weren’t talking to each other.
So Figma growth PM Ezra Mechaber wrote a proposal. The company’s chief product officer saw it, called for volunteers, and a small team built what became PM OS: a repo of 31 AI skills that 60+ PMs now use daily. That name is a joke that stuck, because the interface borrows from Windows XP, complete with a taskbar.
The engine underneath is the same technology that powers Claude skills: each skill is basically a written instruction set that teaches the AI one job. One skill pulls experiment results. Another summarizes stakeholder feedback. Another drafts a launch brief in the company’s format. Stack them, and the AI starts acting less like a chatbot and more like an assistant that knows your job.
Why it caught on at Figma
Plenty of internal AI tools die within a month. PM OS survived because it solved three specific problems, and each one maps to something you can copy.
First, it killed context hunting. Instead of opening six apps to reconstruct what happened on a project, a PM asks one question and the system pulls from connected sources. Second, it standardized the boring stuff. Status updates, launch docs, and experiment readouts follow the same format every time, because the skills encode the format. Third, and this is the big one, it connected to real tools rather than guessing. Skills call the company’s experiment platform directly and read the actual numbers, instead of inventing plausible ones.
That last point matters more than people realize. An AI that writes a status update from vibes is a liability. An AI that reads your experiment data through a connector and cites the real results is an employee.
The parts you can steal for your own job
You don’t need to work at Figma to use the same pattern. The PM OS structure breaks down into four pieces, and all of them translate to any knowledge-work job:
- A skills library: written instructions for your repetitive tasks, each one teaching the AI a single job
- Tool connections: the AI reads your real data (calendar, docs, spreadsheets) instead of guessing
- Formats that never drift: every output lands in the same shape, so nobody has to fix it
- A human checkpoint: one moment where you review before anything ships
Notice what’s missing. There’s no automation running wild overnight, no agent making decisions unsupervised. That design keeps judgment with the human and hands over only the typing.
Build your own AI skills system in Claude
Here’s the practical version. You can build a personal AI skills system in Claude this week, and the whole thing is just markdown files.
- Pick your three most repetitive tasks. Be honest about which ones you do weekly. For most people that’s status updates, meeting summaries, and some kind of report.
- Create a skill for each. In Claude, enable the skills feature and create a new skill folder with a short instruction file. Write it like you’d brief a new intern: what the task is, what format you want, what to never do.
- Include a real example. Paste in your last status update, the one your boss approved. Skills learn format from examples better than from rules.
- Connect a data source if you can. Claude can read files you attach. If your numbers live in a spreadsheet, attach it and tell the skill to pull from it rather than invent figures.
- Run it on last week’s work first. Test against something already done, so you can check the output against reality before you trust it live.
Your first skill will be clunky. The third one takes ten minutes, because you’ll have learned the trick: specific instructions, one job per skill, real examples attached. If your skills start feeling stale or drift from how you actually work, we’ve covered how to fix stale Claude skills with NotebookLM.
A real skill example (steal this format)
To make this concrete, here’s what a finished skill looks like for a recruiter. The instruction file is six lines:
- Job: turn raw interview notes into a candidate summary
- Format: 150 words max, three sections (fit, concerns, next step)
- Source: only use the notes I attach, never add assumptions
- Tone: plain English, no buzzwords
- Example: see attached summary of my last approved writeup
- Never: mention age, gender, or any protected characteristic
Six lines, one job, zero creativity required. Swap “candidate summary” for “listing description,” “lesson plan,” or “weekly client update” and it works for any job that produces documents. That’s the whole trick behind AI skills, and it’s why the Figma team ended up with 31 of them instead of one giant prompt.
There’s a sneaky benefit here too. Written skills double as documentation for your own job. When you go on vacation, the person covering for you reads your skill files and instantly knows how every recurring task works. Managers at Figma noticed this: onboarding new PMs got faster because the skills captured tribal knowledge that used to live only in people’s heads.
One caution from experience: don’t build twenty skills on day one. Three working skills beat twenty half-finished ones, because half-finished skills are the ones that make you stop trusting the system.
The one checkpoint rule
Here’s the discipline that separates useful systems from AI slop, and it’s the same rule the Figma team uses in AutoXP, their agent that runs growth experiments: one human checkpoint before anything goes out. The agent does the work, you review the result, then it ships. Ezra demoed the whole setup, AutoXP included, on Lenny’s Newsletter, and it’s worth a listen if you want to see the real thing in action.
AutoXP can run an entire growth experiment loop, but a person approves the key decision. That’s not a limitation they’ll remove someday. It’s the design. That checkpoint is what lets everyone else move fast, because nobody’s worried about what the AI might have sent overnight.
If you want to go deeper on customizing AI tools to your workflow, our guide to Claude Code mods covers the same philosophy from the coding side. And if you’re wondering which abilities are actually worth building right now, we made a list of AI skills that stopped being worth learning, so you don’t waste a week on the wrong thing.
If you take one thing from PM OS, take that. Automate the typing, keep the judgment. Your first skill is 15 minutes away, and your future self, the one who isn’t reformatting status updates at 6pm on a Friday, will send thanks.