How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)

Ezra Mechaber says he arrived at Figma expecting everyone to be heavily invested in AI, partly because online posts about AI’s future made him feel behind, but found that almost nobody outside engineering was using it for real work. After Claude Code enabled him to build command-line tools, he and fellow PMs Jinnan and Lawrence responded to a challenge from then-CPO Yuki: how could Figma get its PMs “AI pilled”? They turned Ezra’s personal experiments into a reusable system. His goal was not to automate every responsibility, but to ensure neglected PM jobs were done at a minimum acceptable level.

Ezra describes PM work as too many jobs competing for attention, forcing people to decide which tasks they do poorly or skip. AI could “raise the floor” by making baseline work reliable while also “taking off the ceiling” for ambitious users. His examples include automatically looping in Katie from support operations at the right point in a launch, producing good-enough meeting notes so others are not blocked, finding the right code owner in Figma’s roughly 2,000-person organization, and generating detailed specifications that explain how the codebase works. The host highlights that this foundational context enables more strategic workflows.

Ezra demos PM OS as a lightweight repository rather than a conventional app. It contains an agents.md file defining behavior, 31 skills that specify when to activate and what to do, command hooks, and folder structures. On startup it reads the PM’s relationships and memory; one skill finds team members and owners so documents automatically list the relevant designer, engineering manager, and skip-levels. The parent Git repository keeps 60–70 PMs roughly synchronized, while each user has a private workspace sub-repository for writing changes back. A boot hook checks for new PM OS versions.

Ezra says Figma’s Slack is so unruly that he archives about 10 channels a day, so he uses the “What’s Up” skill roughly six times daily. It reads selected channels and full threads, combines Slack and project information, and identifies what Ezra owes people, which decisions remain outstanding, and what deserves attention. He also has a version that reviews his live experiments, although he jokes that he does not peek at directional results. The skill is intentionally tuned to Figma’s organizational structure and helps him operate while juggling too many meetings and projects.

The initial PM OS team worked in a small Slack channel, submitting as many pull requests as possible and creating a “Cambrian explosion” of roughly 1,000 skills before reducing the collection. At a spring offsite, every Figma PM installed the system, used it, modified it, and criticized it; some current skills came from that process, while others disappeared. Ezra says he now removes instructions that better models no longer need. There is no usage telemetry because he did not want colleagues to feel monitored and lacked time to instrument the system, but PMs, engineering managers, and data scientists supplied feedback through public channels and forks.

Ezra’s favorite skills include Build Artifact, which converts a request into a simple HTML artifact using reusable templates and CSS, then publishes it through Figma’s internal hosting tool. Figma’s MCP integration is central to another workflow: a PM-built FigJam builder rapidly creates workshop templates while incorporating design preferences and accumulated “taste.” Ezra also relies on Coaching, which reviews a person’s career ladder, position, and feedback before advising him on a management problem. It may tell him to stop worrying, improve communication with someone he previously misunderstood, or continue working on a specific development area.

PM OS began as a Claude Code terminal interface, leading debates about which terminal app was easiest for people coming from graphical tools. The team then built a local web front end with Markdown formatting, mobile support, and playful ASCII art so PMs could write without staring at monospace text. Ezra briefly built a native version, but now uses Codex almost exclusively because browser use has become strong. Codex can book his desk on a schedule, bash bugs, and run predefined test paths from FigJam boards. In one pass it found that 1,000 emojis broke a text box and that a team-name edge case required “an” rather than “a” in a seat-request flow.

As a growth PM, Ezra calls wasting unallocated traffic the biggest sin, so he built AutoXP to act like a continuously running growth PM. Three Cursor automations communicate in an AutoXP Slack channel: a data-science agent examines funnel drop-off and mix shift, a UX-writing agent proposes copy, and an engineering agent selects and runs a pull request. Someone can request an experiment on a surface such as the settings page, or AutoXP can search for one on a loop. Humans review the data findings, hypothesis, writing, and code once before launch; afterward, results trigger analysis and the next iteration. Five experiments were running, and the first found a major win.

Ezra describes Jev as a fast, inexpensive model that answers constrained questions with a score, classification, or choice. He built a writing editor that calls Jev whenever typing stops, looking for unfinished blanks, unsupported claims, or questions that need answering. When the writer enters “we spoke with X writers and they said,” Jev recognizes the incomplete thought and searches a connected library for relevant interview quotes. Ezra says this combines pattern detection with real-time retrieval rather than merely flagging placeholders. He was so excited by Jev after its launch that he talked about it in nearly every meeting for two weeks.

In the lightning round, Ezra says Instinct’s proactive iMessage interface initially felt magical because it remembered context and could buy items such as paper towels. His current practical use is texting it a pothole location while biking with his children in Brooklyn; it remembers the reporting process and submits the request to 311. He mostly moved to Dots while becoming “fully Codex built,” although Claire says she is trying to build Dots and uses Grok for work. When AI fails, Ezra avoids yelling: he treats it like a friend who is messing up, partly because bad results often reflect poisoned context or poor instructions. His remedy is usually to start a fresh conversation.