I built the patch

I am, on a good day, an idiot

Not the dangerous kind. The useful kind. The kind who notices a hole in a system, cannot leave it alone, and will spend a Saturday trying to patch it with whatever is to hand.

Here is the timeline, because the timeline is the joke.

On 13 August 2026, Google started rolling out Take notes for in-person meetings. Phone on the table. No video call. Gemini writes a Doc: summary, actions, full transcript. Lands in Drive. Emails you the link.

Around the 15th, I ordered an Omi. A little wearable microphone. I had been looking for a way to capture rooms the same way Google Meet already captured calls, and I could not find one that actually filed the result where the rest of the work lives. So I bought a gadget to plug the hole.

This morning, 29 August, I paid for the yearly subscription on it. Properly. Not testing. Paying.

This afternoon I came home with a long recording from a conversation with my mum — she sits on a local committee, I had gone up to help her get files into the right place — and I spent a couple of hours trying to build a better patch. A web app that could take a three-hour audio file, chew through it, and spit out something you could actually use. I already had a smaller version of this on the site. It is fine under about an hour. The recordings I care about are not an hour.

Then I found the button. It had been sitting in Google Meet on my phone since the 13th. I had bought the wearable two days after they shipped it. I had paid for a year of it this morning. I had spent the afternoon reinventing it.

I stood in the kitchen, talked into the phone for thirteen minutes, and three minutes later there was a Doc in Drive.

That is the idiot story. I am going to leave the Omi there. It is a different article. This one is about why I was patching the hole in the first place.

The hole was never “transcription”

If you have been on a proper Google Meet call in the last year, you already know the good version. The meeting does not evaporate. You get the recording, the transcript with speakers, the timestamps, screenshots of what was on screen, a summary, a list of who said they would do what. It is not minutes typed up on Sunday. It is the meeting, filed.

A room does not do that. A committee around a table, a kitchen, a walk. The words happen, everyone nods, and then the only copy of what was agreed is five people’s slightly different memories. Next month you argue about what you decided last month. The AI you paid for cannot help, because it was not in the room, and nothing from the room ever landed in the filing cabinet.

That is the hole. Not “I wish I had a transcript.” The hole is: the spoken work never becomes files, so the digital staff I wrote about in the other post are still sitting at an empty desk.

I have been trying to patch that for a while. A Canvas app on this site for a recording you already have. A wearable for when you are not at a laptop. A custom pipeline this afternoon for files that are too long for the first thing. All of those are patches. Some of them are still useful — if the recording already exists as a file, you still need something that will eat the file. There is a Meeting Notes from Audio tool for that. Use it.

For being in the room, the platform just ate the hole. Phone out. Take notes. Stop. Doc.

I should have looked for the button before I paid for the gadget. I did not. That is the idiot part. The useful part is what the Doc is for.

Decimal places, said out loud

In the digital-staff post I used a slightly annoying metaphor, and I am going to use it again because it is still the right one.

A language model with no context is a whole number. Ask it to write a letter and you get 1. Generic. Fine. Forgettable. Give it the real files — the decisions, the prices, the “we already tried that in March” — and the answer becomes 1.25. Then 1.253. The extra digits are not decoration. They are the difference between a temp with amnesia and someone who has been in the business for a year.

Most of those digits never get typed. They get said, in rooms, by people who are trying to stay present and do not want to take minutes. So they never arrive. The AI stays on 1.0 forever, and you conclude that AI is rubbish, and you are half right. It is rubbish at you, because it has never met you.

This is the bit people want explained with a magic word like “semantic,” and I am going to refuse.

A spreadsheet is extremely good at maths. An empty spreadsheet is still extremely good at maths. It will not notice that March is always expensive, because there are no numbers in it. Put three years of invoices in the cells and you do not have to prompt it to add them up. You also do not have to prompt it, every time, to “please notice the pattern.” The noticing is what you get for having put the numbers where maths can reach them.

Language models are that trick with words instead of numbers. They are not a person in a box. They are engines that do maths on letters. That sounds like an insult until you realise what it implies: they can only do the clever noticing on text they can actually see.

Once the meetings are in the cabinet — not a pile of voice notes in an app, not a recording on a phone, a Doc, on the organisation’s Drive — you stop having to write the prompt that says “please remember what we decided about the hall.” You will not remember to write that prompt anyway. The useful behaviour is the stuff you would never think to ask for. The unfinished action from January that is still unfinished in April, even though nobody used the same sentence twice. The quiet contradiction between what the committee said and what the poster went out saying. The fact that the same name keeps being volunteered for everything.

You do not get that from being good at prompting. You get it from the letters being in one place, in a shape the maths can chew. That is the whole trick. It is not a god. It is a filing cabinet that can read.

What the button actually does (and what it doesn’t)

I used it this afternoon, so this is not a press release.

You open Google Meet. You do not start a call. You tap Take notes, put the phone where it can hear, and talk. Pause if you need to. Stop when you are done. A Doc shows up in Drive with a summary, action items, and a transcript. You get an email with the link. I talked for thirteen minutes in a kitchen. The Doc was there in about three.

It is not the full Meet artefact. There is no video, because there is no video. There are not separate audio feeds for each person, so it will not label speakers the way a proper call does. Google’s own write-up says you want a decent length of conversation — they talk about fifteen minutes — and one language at a time. You need the right subscription, and if you are on a work account your admin has to have it switched on.

It is also not useful if the Doc lands on the wrong Drive. A personal Gmail copy of a committee meeting is not the committee’s memory. It is your hobby file. I will not turn this into a tutorial about ownership and folders. I will just say: the button is the easy part. Putting the file where the organisation actually lives is the bit people skip, and then they wonder why nothing compounds.

Online meetings still win if you can have them online. The recording, the screen, the speakers, the timestamps — that is the version you want when the meeting can be a call. The new button is for the other ninety percent of working life, the bit that happens around a table.

So what do you do on Monday

If you are in the room: phone out, Take notes, stop, file the Doc in the right Drive. That is the patch I was going to build, and you do not need me for it.

If you already have a recording sitting on a computer — last month’s AGM, a voice note, a three-hour conversation you made because you thought you were going to need it — the button will not eat that file. That is what Meeting Notes from Audio is for. It was a patch for a hole. Part of the hole is gone. That part is not.

If you bought a wearable in the last fortnight because you also could not find the in-person version: I am not going to tell you that you are an idiot. I paid for a year of one this morning. I will write about that separately, once I know what I actually think. I found the native button sixteen days late. That is the fact. It does not make the gadget useless. It does mean I should have looked at Meet before I opened my wallet.

And if you have been using AI like a search box that forgets you every morning, do not start with gadgets. Start with the cabinet. The other post is the map. This post is one more way the map gets filled in without you typing.

I spent today being the person who builds the patch. Sometimes that is the job. Sometimes the platform has already done it, and the job is to notice, put the file in the right place, and go to the gym tomorrow instead.

— Kris Wilkinson, Workington