Peter Yang interviews Sam Stephenson, co-founder of the AI meeting-notes company Granola, about using meeting context before and after calls—and how Granola’s team uses agents internally. 33 minutes on Peter Yang’s channel.

Before, during, and after a meeting

  • Pre-meeting briefs are Stephenson’s most important use case: Granola retrieves just enough context about a candidate or unfamiliar person to prevent an underprepared conversation
  • Automatic notes cover roughly 95% of his needs; structured templates matter more for repeatable interviews, sales calls, and other meetings with a fixed rubric
  • “What did I miss?” is Granola’s most-used recipe—useful after multitasking, zoning out, or when reviewing questions and signals that went unnoticed
  • Teams use post-call analysis to ask what they should have asked, how they could have gone deeper, and what to improve on the next sales call

Design for calm, not another dashboard

  • Granola’s primary design goal is an emotion: calm in the chaotic minutes between back-to-back meetings
  • The transcript is deliberately hidden despite complaints; Stephenson would rather preserve attention on the other person than turn the meeting into transcript-watching
  • His metaphor is handwriting: Granola should sit at the edge of the experience and support the user, not become the main character
  • A useful AI assistant should lower blood pressure, not convert every conversation into a louder stack of alerts and tasks

Meetings as organizational memory

  • Repeated use accumulates unusually detailed context: people, projects, decisions, customer signals, and work that may be drifting off course
  • Stephenson turns conversations into first drafts of slide decks, job descriptions, product announcements, and other documents
  • Shared team spaces let Granola answer questions across opted-in calls—for example, identifying themes from the sales team’s recent customer conversations
  • API and MCP access already let users combine meeting context with email, documents, and other systems for recurring reports

When agents skip the interface

  • Stephenson admits a “pang of sadness” at the idea that agents may use Granola’s API without showing its carefully designed UI
  • Enterprise reality forced the company to accept this: meeting context must connect to internal tools, coding agents, CRM, Slack, and company-specific workflows
  • The product’s durable role may be capturing and organizing high-quality context—even when the next action happens somewhere else
  • Granola plans to learn from common MCP workflows and productize the ones users repeatedly build for themselves

AI-native work without the 996 grind

  • Stephenson rejects the idea that AI-native means working nonstop; he argues that life outside work and time away improve the thinking that determines what to build
  • Coding agents can become a slot machine—“one more prompt”—while switching among several agent threads can be more exhausting than switching meetings
  • The hard product problem is not extracting three tasks from every call; it is helping with the resulting pile without making the user feel more overwhelmed
  • Faster implementation makes judgment more important: the scarce skill is choosing something worth making, not staying busy generating output

How Granola keeps agent-built software from becoming slop

  • Everyone can build internal tools, and a surprising number of non-engineers’ changes reach production
  • A strong design system—tokens, reusable components, and documentation agents can follow—makes routine agent-generated UI look coherent without a designer blocking every change
  • Designers now move between whiteboards, Figma, and live code; Stephenson estimates his own Figma use has fallen to about 20%
  • Static mocks still help with copy-heavy linear flows, but interactive features are tested in code with real personal data because their value depends on timing and context
  • Teams ship ugly end-to-end prototypes internally within a week or two, then dogfood them until the pain creates pressure to improve the product

Nacho, Granola’s internal first responder

  • Nacho is a shared internal agent connected to Slack, Granola, Amplitude, logs, code, and website analytics
  • It is the team’s first stop for operational questions: query the data, inspect logs, compare the failure against code, and explain a likely cause
  • Nacho can hand an investigation to a Cursor agent, turning diagnosis into a proposed pull request
  • Nobody formally owns it; an engineer began it as a side project, and employees keep improving it because every improvement benefits the whole company

Shipping is the start, not the finish

  • Stephenson’s warning to AI builders: easier implementation raises the value of craft rather than eliminating it
  • Ship rough versions early to coworkers, friends, and beta users—but do not confuse deployment with success
  • Most builders stop at version one; the real standard is whether somebody receives useful value every day
  • Granola’s target is not “we shipped it,” but “this reliably works for people”

“The easier it gets to build stuff, the more it matters to really sweat the details and go the extra mile to make something that’s actually going to be useful to somebody.” — Sam Stephenson