Meeting Transcript to Professional Minutes—The Claude Code Flow I Use Almost Daily

Some tech titan said the other day that Mythos 5 is already powerful enough to go build a 3D-engine game. But for most of the people watching this, that has absolutely nothing to do with us. We won’t touch it at work or at home, and beyond the spectacle it does nothing for our efficiency. Better to spend the time learning the AI skills we can use at work and in life.

Today I’m sharing the thing I use most in my daily work: turning a transcript of a meeting recording into minutes that everyone who attended finds genuinely useful. My online meetings with clients involve trade secrets and aren’t fit to show, so I’ll use the first chapter of Demi-Gods and Semi-Devils to illustrate the whole process.

Step one: correct it first, so the minutes aren’t “recorded but useless”

Paste the transcript exported from your online meeting software or recording device into Claude Code. The prompt goes roughly like this: this is the speech-to-text transcript of my online meeting with a client today; the transcription may contain recognition errors or dropped information; read the whole thing, decide which parts need correcting, list them for me to confirm, and get me to a complete, reasonably accurate transcript.

Why this step is indispensable: speech-to-text accuracy is only about 60%-70%, and when the room is noisy or the mic is far away you get a lot of errors. Information doesn’t go missing so much as get mangled—noise makes the model hear one word as a homophone of another. One correction pass usually repairs the content to 90% accuracy or better.

A lot of people skip this step when they write up minutes, which is why their minutes end up with hallucinations in them—recorded but useless. We take minutes to move work forward; being able to send them out without editing the content is itself a real time saver. But that 20%-30% error rate is exactly why we end up recording a mountain of meaningless meeting notes.

In my test, the Demi-Gods and Semi-Devils text I used looked like a PDF run through OCR, and it had a fair number of errors: one character description came out garbled, and Claude caught every one of them from context. The classic case in day-to-day meetings is names—the same person comes out as two different people. Once you’ve confirmed everything, close with one line: “Your corrections are all right—generate the final accurate transcript accordingly.”

Step two: generate web minutes structured to the meeting type

Next, have it work out the meeting type from the transcript, write the minutes to a structure that fits that type, and render them as a web page in the PPVI light design style—that’s the skill I’ve covered in other videos.

There’s a lot of room for customization here: if your company has a specific design style, just ask for it. If you’re serving a client, have the page use the client’s brand colors—Coca-Cola being the classic red and black. For less well-known clients, give it the company website and let it find the brand colors itself. Same with the structure: what framework your company’s usual client or consulting meetings call for, what goes in each section and how long it should run, what logic to order it by. You can feed it templates up front. Even a vague brief is enough for it to work from.

The minutes it generates contain the core conclusions, the parties involved, a timeline, the discussion of each agenda item, risks and open items, and next-step action items—and sometimes I also have it analyze the need behind the client’s stated need. For business consulting, any business canvas or business unit model we worked through with the client during the meeting also gets rendered graphically.

A side note: it recognized that what I gave it was the original text of the first chapter of Jin Yong’s Demi-Gods and Semi-Devils. Claude’s safety guardrails are still strong—outputting copyrighted original text is prohibited, and it can only generate summary or transformative content.

Step three: a one-page PDF, sent straight to everyone who attended

I was happy with the web minutes, except at the end it generated a “meeting record quality note,” which I don’t need. The next step brings in another skill: strip that section out and save the page as a one-page PDF. That’s onepage-pdf, which I covered in another video—it turns the page into a PDF you can send over a chat app to everyone who attended. This is my most-used feature by far; I use it almost every day.

Wrapping up: you really don’t need note-management software

I have Claude Code save pages and PDFs like these into one fixed folder on my D drive, and I never organize them—there are far too many to organize. But Claude Code generated every one of them, so it knows the directory structure. When I’m working inside some project, it can pull the right document up as context in seconds and help me push the project forward—a deck, a consulting report for a client, the next round of research, whatever I need next.

No note-management software can keep up with this. Letting Claude Code manage this content itself is the simpler, more practical way.

Source: EP0038_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] Some internet big shot said in the last couple of days that the Mythos 5 model is already powerful enough to build a game on a 3D engine. But for the vast majority of you in front of the screen, I’d say this has absolutely nothing to do with us. We don’t actually use it in our work or our lives, right? Unless you just want the efficiency bragging rights, it’s a sideshow. We’d be better off spending that time actually learning the AI skills we can use in our work and our lives. So today I’m going to share the thing I use most in my own day-to-day work,

[00:25] which is taking a meeting transcript—a transcript converted from a recording—and turning it into meeting minutes that every attendee finds genuinely easy to use. Let’s jump straight into it. Normally my online meetings with clients are consulting sessions involving trade secrets, so I can’t really show you those here. What I’m using as today’s example is the text of chapter one of the novel Demi-Gods and Semi-Devils, which I found online. I’ll use that content to walk through the whole process. First we

[00:51] take the meeting transcript we’ve got and right-click paste it into Claude Code. Our prompt goes roughly like this: The text I’m giving you now is the speech-to-text transcript of an online meeting I had with a client today. During transcription there may be recognition errors or lost information in it. I need you to read the whole thing and work out which parts need fixing.

[01:17] Then I need you to list those spots out and confirm them with me, so that in the end we get a complete, reasonably accurate transcript. So step one is handing it the transcript we exported out of the online meeting software, or the transcript exported off the recorder. Honestly, once a recording gets turned into text you only get about 60 or 70% accuracy. Sometimes when the recording environment is noisy,

[01:42] or the mic is far away, there are quite a lot of errors in it. It’s not really lost information—it’s that with the noise it hears a word as a homophone, or just gets it flat-out wrong. So with a correction pass like this, we can usually repair the content up to 90% accuracy or even higher. And that’s the foundation for making meeting minutes—

[02:08] the foundation. A lot of people skip this step, and the result is minutes with hallucinated bits in them, or minutes that might as well not have been taken. You know what I mean? The whole point of taking them is to move the work forward—I shouldn’t have to edit the content before I can send it out to everyone. That right there is where the time savings are. But because of that 20%, 30% error rate, we end up cranking out piles of meeting notes that are just useless.

[02:33] So I’ll, I’ll speed this part up while it produces the output. Okay, you can see it’s already identified all the errors. The copy of Demi-Gods and Semi-Devils I found was probably a version OCR’d out of a PDF, and I see quite a few errors in it. Look—“a middle-aged fellow” got recognized as “a mid-year fellow,” which is an error in the novel file itself, the file I gave it. And look, it caught all of them from context. In our everyday

[02:59] meetings, you get recognition errors on how people are addressed—often the same person’s name comes out as two different people. So you can see it’s pretty much listed out every one of these errors. If there’s nothing wrong with them, the next prompt we give it is: All the corrections you’ve proposed are right. Generate a final, accurate transcript following your corrections.

[03:24] Then, then next we need you to work out what kind of meeting this was from the content of the transcript. We need to generate minutes for this meeting using a structure that fits it. And please generate the page in the PPVI light design style. So in this process, as you can all see, I’m actually using a

[03:49] skill—the PPVI design style I’ve talked about in other videos. It generates the page in that style. But there’s actually room to customize things here. For instance, say I’m a company and I need this page generated in a specific design style—you can ask for that right here too. Say, hypothetically, Coca-Cola is our client. Then right here I’d tell it that the page’s color scheme needs to use the client’s brand colors.

[04:14] Coca-Cola, for example, is classic red and black, right? And if the client here isn’t that well known, we can also give it the client’s website and have it go find the client’s brand colors, then use those in the page it generates next. Oh, it’s already done. Let’s take a look—we actually hit a problem. It figured out that what I gave it was the original text of chapter one of Jin Yong’s Demi-Gods and Semi-Devils. Claude’s safety guardrails are still really strong,

[04:40] and when it comes to reproducing the original text of copyrighted content, it blocks us from outputting it. So here it tells me it can’t generate the transcript the way I asked, but it can generate summary-style or transformative content. So it generated a page. It generated a page, and let’s look at it—here it is, the web version of the meeting minutes, the thing you’d normally see. It has the meeting title, and then based on the meeting type it

[05:06] builds a structure like this. And we can actually hand it that structure ourselves—say, for our company’s typical client meetings or consulting meetings, what structure it should use, what goes in each section, how long each one should be, how it should be organized. You can give it a template like that up front. Even a vague request is enough for it to pull something together.

[05:31] So in here you get the core conclusions, plus the parties involved. Sometimes I also have it analyze the need behind the client’s stated need. Then there’s the timeline, how the agenda items unfolded, risks and open items, next-step action items. What else can you put in there? If it’s a commercial consulting session and we’ve already worked through the client’s Lean Canvas or their business unit model during the conversation,

[05:57] that can go in here too, laid out as a diagram. At the very end it also generated a note on the quality of the meeting record, which we don’t actually need. So what I usually do next is bring in another skill: I’m really happy with this web version of the minutes. That last section, the quality analysis of the minutes, isn’t necessary, so please take that part out, and save this page as a

[06:24] single-page PDF, so I can send it over a chat app to both sides who were in the meeting. Doing it this way calls up the onepage-pdf skill I made and mentioned in another video, and it turns this page into a PDF I can send to everyone who was in the meeting. This is genuinely the thing I use most, most often—

[06:50] a feature I use almost every day. These pages and PDFs, I have Claude Code save them into one fixed folder on my D drive. And they don’t need organizing, because there are too many to organize. But since every one of them was generated by Claude Code, it knows the directory structure too. So when I’m calling on it inside some project, it can pull that document up extremely fast as context to help me push the project further along—

[07:16] whether that’s putting together a PPT, or turning out more for the client, consulting reports, or the material for the next round of research. It’s incredibly handy. You don’t need note-management software at all—you could never keep up with it anyway. So letting Claude Code manage this stuff itself is the simpler, more practical way. That’s it for today. See you tomorrow, bye.