Source: raw/The_AI_chief_of_staff_that_this_Figma_PM_uses_all_day.md

Creator: Claire Vo, How I AI, with Ezra McCabe, a growth product manager at Figma (the captions mishear Figma as “Sigma” once) · URL: https://www.youtube.com/watch?v=ztt8irafLMQ · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-05). Sponsors: Optimizely, Jira Product Discovery.

McCabe walks through three things. PMOS is a repository of 31 skills that Figma’s 60 to 70 product managers install as a “chief of staff”. He also has Codex run bug bashes in the browser from a FigJam test plan, and his team built Auto XP, which proposes, builds and analyses growth experiments from a Slack channel. He ends with a writing prototype that calls Jev, a cheap decision model, every time he stops typing. His stated goal: “be lazy in the right places”, and get PMs to a minimum acceptable level on the jobs they currently do badly or skip.

Key Takeaways

PMOS: a skills repo as a PM’s chief of staff

  • Why. At Figma he found almost nobody outside engineering using AI for real work. The CPO at the time, Yuki, asked what it would take to get every PM “AI-pilled”. McCabe and two other PMs turned his own Claude Code habits into something reusable.
  • What is in the repo. It is “a really, really basic repo”:
    • an agent.md that says how to behave (“you’re the PM’s chief of staff, you know their work”);
    • a memory file and a profile of who the PM is and who they work with, read at session start;
    • skills that say when to fire and what to do;
    • command hooks, for example on boot check for a new PMOS version, to keep 60 to 70 PMs roughly in sync;
    • a hidden dot-folder base template that onboarding copies into the PM’s own workspace, so they cannot break the shared context;
    • two git repos: the shared parent everyone downloads, and a private workspace sub-repo each PM can write back to.
  • How the skills were chosen. Three PMs in a small Slack channel sent “as many PRs as we could” (a “Cambrian explosion of like a thousand skills”), then cut them down. An offsite with every Figma PM had everyone install it, use it and “roast it”. As models improved he has been removing skills that no longer help. No telemetry, by choice: he did not want anyone to feel watched, and the PMs gave plenty of feedback in a public channel. Engineering managers adopted it; the data scientists run a fork.
  • Skills he singles out:
    • Who works on what. Figma has about 2,000 people and several org-chart tools that do not connect. The skill knows your designer, engineering manager and skip-levels and adds them as reviewers when it writes a doc.
    • “What’s up”, run about six times a day. It pulls his relevant Slack channels and projects, reads whole threads, and lists who he owes things to, which decisions are outstanding, and the state of his live experiments. He archives about 10 Slack channels a day.
    • Build artifact. Turns material into a small HTML page from his saved CSS templates and publishes it to an internal hosting tool.
    • FigJam builder. Built by another PM on the Figma MCP server; produces workshop templates.
    • Coaching. Given the career ladder and the feedback he has received, it does the first pass of coaching before he takes a problem to his manager, including reminders such as last time this person and you simply misunderstood each other.
  • Interface history. Claude Code in the terminal, then a local web front end (ASCII-art styling, markdown rendering, works on mobile) because PMs write a lot and “it sucks to read monospace font all the time”, then a local native build. He is now “basically in Codex 100% of the time”, mainly for browser use. A scheduled task books his desk in the New York office one day ahead.

Bug bashes run by Codex in the browser

  • Setup. A Codex thread that “just lives here all the time”. For each launch, a FigJam file lists the test account to use, the path to follow, the actions to take and what to look for. He takes one test account and gives the agent (on OpenAI’s Astra model) another.
  • Loop. It runs every scenario in the browser and writes what it did back onto the FigJam board. He is still tuning it so its comments are not annoying.
  • What it found in the past month. A text box that broke when given a thousand emoji, and a seat-request message (“a [team name] seat”) that needed “an” for team names starting with A, which earlier human bug bashes had missed because it depends on the team name. “Astra is just much much more diligent about following directions than I am.”

Auto XP: an agent that runs growth experiments

  • The premise. “The biggest sin you can commit is wasting traffic”: any traffic not allocated to an experiment is waste. Auto XP runs experiments on surfaces nobody would have spent time on.
  • Mechanics. Three Cursor automations triggered from a dedicated Slack channel, where Auto XP also talks to itself.
    1. A person tees up an experiment (“go find an experiment to run on our settings page… Go crazy”), or it runs on a loop.
    2. A data-scientist agent looks for drop-off, the funnel and mix shift.
    3. A UX-writing agent proposes copy (front-end design is being added).
    4. An engineer agent opens a PR.
    5. Each step posts into the same Slack thread.
    6. When the experiment ends, a webhook sends it back to analyse the results, using his written instructions on when to explore further, when to exploit, and how to call a win or a loss.
  • One human touch. People review the data-science findings, hypotheses, copy and code together, once, before launch. After that they only see the result, by which time it has started on the next round.
  • Status. One engineer, Maryanne, runs it, with five live experiments “that I have not taken a look at”. McCabe says it “found a huge win on its very first experiment”, without numbers, and puts it at “great” in theory, “almost there” in practice. He says people elsewhere (the captions are unclear on where) had built similar loops, which inspired it.
  • Vo, who founded an experimentation startup about 13 years ago, calls it what that company was trying to do.

A writing check that fires when you stop typing

  • Idea. He found Jev fast and cheap enough to call constantly, unlike the frontier models behind Figma’s design-canvas agent, which feel slow for a co-worker. The prototype was inspired by a writing app he recalls as Marker.
  • Behaviour. Each time he pauses, it calls Jev to find blanks (“TK”, brackets) and claims worth supporting, then searches his library of interviews and offers matching quotes. Typing “we spoke with X writers and they said” pulled interview quotes from the library. The live demo was blocked because the API now sits behind Figma’s firewall.
  • His worry: candidates in product-sense interviews running the same kind of helper against their own career stories.

Other notes

  • Personal agents. He still uses Instinct, which works over iMessage, to report potholes to New York’s 311 service (“Hey Siri, text Instinct there’s a pothole at…”), but has mostly moved to Dots now that he works in Codex.
  • When the agent is stuck. He does not yell. A bad result usually means a poisoned context or a bad question, so start a new conversation. His example: a Codex session moved from local to remote inherited the ask-permission sandbox even with YOLO mode on, and arguing inside that thread did not help.
  • Vo’s counter-habit: set high expectations (“buddy, I know you can cook… I know this is in your capabilities”).

Try It

  1. Start a team skills repo with the boring context first: a who-owns-what skill and a “what’s up” sweep across your Slack channels, owed items and open decisions. Vo’s point: most PM OSes skip this layer.
  2. Separate shared from personal. Keep the shared repo read-only (a hidden base folder) and give each person a private workspace repo, with a boot hook that checks for a newer shared version.
  3. Prune on a schedule. Each time the model changes, remove skills and see if output gets worse.
  4. Run your next bug bash with a browser agent: write the test plan (account, path, actions, what to look for) in one document, give the agent its own test account, and have it write findings back to that document.
  5. Before an Auto-XP-style loop, write down your analysis rules (when to explore, when to exploit, what counts as a win), and keep a single human review of findings, hypothesis, copy and code before launch.
  6. Wire a cheap model to your editor’s pause event to flag “TK” placeholders and unsupported claims, and look them up in your own notes.

Open Questions

  • No figures are given for Auto XP’s “huge win”, and the experiments’ outcomes were not reviewed on air.
  • The writing app he names as inspiration (“Marker”, by someone captioned “Stein Deck”) could not be confirmed from the transcript.
  • Which agent runs inside Figma’s design canvas is not stated.