ExpertsFMHow Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)
How I AI · November 2, 2026
5 Selected Lessons:
case · 2 heuristic · 2 framework · 1
Context-aggregation skill layer before complex workflows
Source
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma) (November 2, 2026) · at 6:35 ▸ · at 9:35 ▸
When this applies
when building any AI agent for a complex organization with scattered information sources
The takeaway
Before building sophisticated agent workflows, invest in a basic context/aggregation layer that answers simple organizational facts reliably.
Why it works
Complex, judgment-heavy agent workflows fail without reliable facts about who owns what, who to loop in, and what's outstanding; building a foundational layer that answers 'who/what/where' first makes every downstream skill more accurate.
One level up
Sophisticated automation is bottlenecked by unglamorous foundational data plumbing, not by model capability.
Where it stops holding
This stops working when the organization is small/simple enough that such context is already trivially known or when the underlying data sources are too inconsistent to aggregate cheaply.
Detail
procedure:
- Identify repetitive context-lookup pain (org charts, who to loop in, outstanding decisions)
- Build a skill that aggregates this across fragmented tools (Slack, docs, org chart tools)
- Have it auto-attach relevant people/context when generating other artifacts (e.g., docs)
- Run this aggregation skill frequently (e.g., several times a day) to keep a human oriented amid overload
rationale: Reliable low-level context retrieval is a prerequisite for trustworthy higher-level agent actions.
In their words
So many of these, like, PMOSs that we've seen skip all this, like, foundational piece, which is in large organization, it is actually really hard to know, like, who should have an opinion on this product?
Move users out of the terminal once they're comfortable
Source
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma) (November 2, 2026) · at 12:00 ▸ · at 17:34 ▸
When this applies
when rolling out a technical tool to non-technical users and they've cleared the initial adoption hurdle
The takeaway
Treat a difficult-but-necessary interface as a temporary bridge to build user comfort, then invest in migrating to a more natural interface once the underlying behavior change has taken hold.
Why it works
The terminal was useful as an onboarding ramp to get comfortable with AI agents, but it is not the natural long-term interface for writing-heavy, non-engineer work; once comfort is established, the interface itself becomes the bottleneck.
One level up
The right interface changes as user sophistication grows; don't anchor to the interface that got you initial traction.
Where it stops holding
This stops working when the initial interface is actually well-suited to the task long-term, not just a stopgap (e.g., when the audience is engineers who prefer terminal).
Detail
rule: Once users adopt a hard interface successfully, prioritize migrating them to an easier one rather than keeping them there.
threshold: Once a critical mass is comfortable in the terminal
rationale: 'it, like, sucks to read monospace font all the time' for writing-heavy work; comfort with the concept matters more than the specific interface.
In their words
We managed to get a bunch of PMs comfortable in the terminal, but, like, surely that's, like, not the right end state.
Pick the lightest-weight agent interface for the actual task size
Source
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma) (November 2, 2026) · at 33:16 ▸
When this applies
when selecting among multiple agent/automation tools of different weight and polish for a given task
The takeaway
Keep lightweight, low-friction agent tools for small spontaneous tasks even after adopting a more powerful primary tool, matching tool weight to task size rather than consolidating everything into one system.
Why it works
A lightweight interface (texting an agent) removes friction for small, infrequent tasks (reporting a pothole, buying paper towels) even if it lacks the depth of a full coding-based agent, so the two coexist for different task sizes rather than one replacing the other.
One level up
Tool portfolio should be diversified by task granularity, not unified for its own sake.
Where it stops holding
This stops working when maintaining multiple tools creates more overhead (context-switching, inconsistent memory) than the friction savings justify.
Detail
rule: Keep a frictionless, low-overhead agent channel for small spontaneous real-world tasks even when a more powerful primary agent handles your main work.
threshold: Tasks that are quick, infrequent, and easy to forget (e.g., reporting a pothole) favor texting-based triggers over app-based workflows.
rationale: 'Hey Siri, text instinct' removes all app-opening friction for tasks you'd otherwise forget.
In their words
Hey, Siri. Text instinct. There's a pothole at, like, 269 Atlantic, and it just remembers, like, what it's supposed to go do and submits it.
Cambrian explosion then whittle-down of skills
Source
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma) (November 2, 2026) · at 12:00 ▸
When this applies
when bootstrapping any new toolkit/feature set where you don't yet know what will be useful
The takeaway
Generate many candidate solutions cheaply and fast, then prune aggressively based on real user feedback rather than trying to predict the right scope upfront.
Why it works
Rapid, low-cost generation of many candidate skills in a tight feedback loop, followed by broad user testing (off-site) and pruning based on real feedback, converges faster on useful functionality than careful upfront design.
One level up
Exploration-then-pruning beats premature optimization when uncertainty about user needs is high.
Where it stops holding
This stops working when generating candidates is expensive or when there's no fast, cheap feedback loop to prune against.
Detail
situation: Three PMs hacking in a Slack channel created roughly 1,000 candidate skills for an internal AI tool.
decision: They ran an all-PM off-site where everyone installed, used, and 'roasted' the tool, generating real feedback.
outcome: Skills were whittled down or removed based on feedback, and further pruned over time as underlying models improved and some skills became unnecessary.
lesson: Mass-generate then prune via live user feedback rather than designing the 'right' feature set in advance.
In their words
We had that sort of, like, Cambrian explosion of, like, a thousand skills and then started to, like, whittle stuff down.
Autonomous bug-bashing agent catches edge cases a human misses
Source
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma) (November 2, 2026) · at 17:34 ▸
When this applies
when doing manual QA/review work that is tedious and judgment-light but requires exhaustive diligence
The takeaway
For tasks that require exhaustive, repetitive coverage rather than judgment, delegate to an agent that will follow the full instruction set diligently rather than relying on a human who tends to shortcut.
Why it works
An agent following explicit step-by-step instructions on a test account doesn't get bored or skip steps the way a human does, so it surfaces rare edge cases (e.g., name-dependent bugs) that depend on exhaustive coverage rather than insight.
One level up
Match task delegation to comparative advantage: humans for judgment, agents for diligence/exhaustiveness.
Where it stops holding
This stops working when the test scenarios aren't well-specified in advance, since the agent can only be as thorough as its given script.
Detail
situation: Team needed thorough bug-bashing of UX changes with complex underlying logic before launch.
decision: Ran a human account alongside an agent ('Astra') account following the same scripted FigJam test plan, having the agent write findings back to the board.
outcome: Agent found edge cases the human bug bashes missed, e.g. a text-box emoji overflow bug and a team-name-dependent article bug ('a' vs 'an').
lesson: Diligent, instruction-following agents can outperform humans on exhaustive repetitive QA, catching nits humans would never notice.
In their words
Astra is just much, much more diligent about following directions than I am. So I will run it through, and I will say, like, keep doing all of these specifically.