The first AI tool whose absence would measurably raise my anxiety level.

Granola is a desktop notetaker — Mac, Windows, iOS — that captures the audio of whatever meeting you are in locally, on your machine, without joining the call as a bot. It transcribes in real time, writes notes to whichever template you point it at, and lets you chat against the resulting transcript afterward to produce derived artifacts. There is no calendar bot, no shared recording link, no “Hi everyone, I'm an AI notetaker.” The tradeoff is honest: it cannot capture audio from people not in the room with you unless your machine can hear them, which is exactly the same condition you already operate under as the human in the meeting.

I do not believe a meeting tool is general. I believe a meeting tool is good at the four or five meeting shapes you actually attend. Mine are below. Yours are different, and Granola will only work if you spend a week noticing yours and writing templates against them.

I write templates as a list of section headings with a one-line instruction underneath. The model fills the body; I edit the body to match how I would have written it. The structure does not change between meetings of the same shape. Stability is the point.

That template runs ~6 times a week. It does not change. When it does change, it is because the meeting shape changed underneath me, which is rare and worth noticing.

The 1:1 engineer template above is the one I run most — six times a week, and the most edited of any I own. Its sibling, the 1:1 manager template, shares the same shape but adds an extra section for org context, which I need exactly twice a week. Both pull their value from being stable: the cost of changing them is high enough that doing so signals the underlying meeting actually changed.

Customer calls come in two shapes that need separate templates. Discovery is the canonical one — use case, friction, the verbatim quote I want my PM to see, the ask I'm carrying back, the follow-up — and it feeds the CRM row directly. Escalation is the same template plus a strict timeline section and a who-owns-it row, which is what turns a tough call into a closeable thread instead of a vague apology.

The remaining four templates each match a specific recurring shape. Hiring screens get rubric signals plus two verbatim quotes plus a recommendation, with no editorializing — the whole point is to keep my opinions visible as opinions. External strategy meetings get a room map at the top, decisions in the middle, parked items at the bottom, in that order. Internal reviews track what shipped, what slipped, the honest reason for the slippage, and the asks for next quarter. Pre-mortems get one section per failure mode and three lines under each: how it happens, how we catch it, what we do.

I have run a meeting tool before. Two, actually. Both got uninstalled inside a month. The difference here is not capture quality — capture quality has been roughly fine across the category for two years. The difference is what happens to the capture after the meeting ends.

I have a quiet belief that the meeting is the cheapest place to put a piece of organizational state. It is also the most lossy place. Granola is interesting because it raises the floor of the lossy place without making the meeting itself worse. The model is not in the room. Nobody is performing for it.

The default outcome of any meeting I attend is a paragraph in something downstream of it. Granola makes that paragraph cheaper to write, harder to skip, and easier to revise the day after.

The single most-asked question I get about Granola is whether it sends a bot. It does not. The single most-useful comparison I can make is against the two architectures that do — the calendar bot, and the screen-share recorder. Each does some things Granola cannot. Granola does the one thing each of those cannot.

I keep finding that the meeting is not what I am paid for; what I am paid for is what happens between the meeting and the next one. That gap used to be where my notes degraded, my memory blurred, and my Friday turned into a re-creation exercise. Granola sits in that gap. Each transcript is the input layer for the four things I had to write anyway.

I shipped a blog bot at WorkOS — Cloudflare Workers + Workflows + Slack + Webflow CMS — that is now used org-wide. Once that pipeline existed, every other Applied AI internal feature I wanted to build became cheaper because the scaffolding was already in place. Granola plugs into that pipeline directly: “draft our perspective on the customer questions from this morning's Acme call” stops being a thirty-minute writing task. The capture is a means to an end, and the end is that the same hour produces four artifacts instead of zero.

I pay for the Business plan ($14/user/month). I had Basic for the first six weeks; I moved up because I wanted unlimited meeting history and the better integrations. The comparison below is current as of May 2026 — confirm on granola.ai for the latest.

The honest math: I spend about an hour a day writing artifacts that derive from meetings — Slack updates, CRM rows, follow-ups, internal docs. Granola does not remove that hour; it shifts where the time lands. The bar below scales to your meeting load. Move it.

These are the chat prompts I run against a finished transcript. None of them are clever — they are the operations I used to do by hand at the end of the day, each refined over twelve months by deleting the parts that did not survive into the doc I actually sent.

The first is the Slack update. After every meeting that warrants one, I run “Write a four-bullet Slack update to my team about this meeting. Lead with what we decided. Use my voice — short, first-person, no jargon.” The output ships within a minute of the call ending, before anyone has to ask what happened. For customer calls the equivalent is the CRM row, generated from a stricter prompt — “Output a single-row update for our CRM with these fields: Use case · Friction · Verbatim quote · Asks · Next step. One line each.” Paste, save, done.

Follow-up emails get a similar treatment, scoped to one recipient and capped at six sentences: “Draft a follow-up to [name] that thanks them, restates the two decisions, and asks for the one thing I committed to ask.” On Fridays I run a different kind of prompt across the whole week's captures — “Across all meetings tagged #1:1 this week, list every commitment I made with an owner of ‘me’ and a due date. Sort by due date.” That one is the closest thing I have to a personal accountability dashboard. Once a quarter I also run a pre-mortem rewrite — “Re-read this pre-mortem. For each failure mode, write the one sentence we would say in the retro if it happened.” Sharper hypothetical, sharper meeting.

The prompts are not the value. The constraint behind each prompt is — knowing the shape of the artifact you want before you ask for it. That part transfers.

Capture quality drops in the cold-room case — a conference room with three people, one phone speaker, one whiteboard. The fix is to be the person near the speakerphone, not to expect magic. The related failure mode is overlapping speakers: three voices on top of each other get collapsed, and the transcript reads like the loudest person was the only person. Granola does not fix the meeting, only the writing about the meeting.

The first week is also a tax. You will not have templates yet, and you will not yet have a sense for which prompts produce which artifacts. By week two it stops being a tax; by week four you stop noticing it's there. Multi-language meetings are a real edge — capture is strong in English, decent in Spanish and French, weaker elsewhere. If half your calendar is bilingual, test it on yours before committing.

The honest stretch is using Granola past meetings — long phone calls with vendors, for example. It works, but the output is rougher because the meeting shape is rougher. The tool is honest about that, and so am I.

Install it before Sunday. Open it once. Click through the onboarding. Do not write a template yet — let the app sit there until Monday. Then run it on your first real meeting using the default template, read the output, and notice what's missing. The point of the first run is not to like the output; it is to discover what you want.

Monday evening, write one template — the shape of the meeting you have the most of. Five sections, a one-line instruction under each. Twenty minutes, no more. Then leave it alone for the rest of the week and use it. By Friday you'll either trust the output or have a specific edit in mind. Either is fine.

After a week, add a second meeting type, then a third. Build your prompt pack one template at a time, anchored to a meeting shape you actually attend. And tell people. The local-audio architecture lets you be silent; that doesn't mean you should be. “I'm taking notes with an AI tool” is one sentence. It costs nothing, and it closes the consent loop so you stop having to think about it.

The first week of running Granola is the only week where it's a tax. By week two you stop noticing it's there, except for the part where you start noticing how much less you're carrying.

I do not run Granola in calls where consent is unclear. Personal capture is one thing; recording someone who has not understood that capture is happening is another, and the local-audio architecture makes it technically silent without making it ethically silent. I tell people. The one-sentence disclosure costs nothing and resolves the awkwardness before it starts.

I do not paste raw transcripts back at people, either. The summary is for me. The transcript is for me. What goes to my manager, a customer, or a partner is something I wrote, informed by the record, not pulled from it verbatim. The line between those two artifacts is the thing that protects the relationship.

And I do not pretend the AI is doing the listening for me. It captures. I still have to think, still have to remember, still have to actually be in the meeting. Granola removed the part of meetings I was bad at; it did not remove the part I have to do.

The card below opens Granola directly. The form further down sends my workflow guide first. Either path works.