I Let Claude Edit My Podcast

A month ago I sat down at a cafe with Daniel Koenawan, a mate from my PwC days, put a camera on the table and hit record. The idea was simple: two Indonesian techies having a yarn about how technology has changed the way we work. The conversation was great. The editing was the problem.

I've done one podcast before, and editing it nearly killed the idea. I'm not an editor. Finding what to cut meant watching the whole thing two or three times: once to know what's there, again to catch the private stuff and the names, then again to stitch it into something that isn't crap. This time I opened the Insta360 desktop app, started scrubbing through clip by clip, and a few minutes in I could tell it was going to be the same slog.

So I turned it into a test. I'd been wondering whether AI could genuinely do this kind of work well, and I wanted to prove it on something real of my own before ever suggesting it at work. This podcast was the perfect guinea pig.

The episode's at the bottom if you want to jump straight to it.

Two Things That Made Me Try It

For a while now I've been using Claude to make how-to videos for the apps I build. I take screenshots as I click through, no narration. Claude works out what I'm doing from the screens, and because it has access to the code base, it knows what each screen is actually for. It writes the steps, stitches the frames into a video and times it all. I add background music from Suno, another AI that makes music from a prompt. The first time I saw the result, I was flabbergasted.

The second thing was my accent. I'll admit I'm a bit fixated on whether I sound Kiwi, since I've lived here longer than I lived in Indonesia. When I asked ChatGPT, it was honest: it can't hear an accent in a transcript, only the words. Give it the actual audio, though, and it'll have a go. Its verdict was about 70% New Zealand, 30% Indonesian. Not science, but I'll take it.

Sitting in the Insta360 app, those two clicked together. If Claude can stitch frames and audio, and AI can work with an audio file, could it transcribe my podcast with timestamps and tell me where to cut? Could it work with actual MP4 footage? So I asked Claude, more or less: am I asking too much here? Its answer was that it could do the stitching too. That's when this went from "save me some scrubbing" to "do the whole first pass".

Why I'm Doing a Podcast at All

Us guys are told we need to talk more. We catch up with mates maybe once a year, and only if someone else organises it. A podcast is my excuse to fix that: grab a coffee with a friend, press record, have a proper conversation.

It's also practice. I used to do Toastmasters, and one bit of advice stuck with me: post it on YouTube. That's skin in the game. It's public, it's a bit uncomfortable, and that's exactly why you get better, including at hearing your own voice played back. If the editing takes hours, though, I'll stop doing it. So the editing had to get easier.

An SD Card and a Lot of Questions

The footage was on an Insta360 Ace Pro: seven clips, because we started and stopped a few times, 4K at 60 frames per second. About 83 GB. Claude Desktop can link to your computer, so I connected a folder and plugged in the SD card. My first worry: what if it deletes my footage? Deleting files on my machine is blocked unless I approve it, and most SD cards have a lock switch on the side anyway. Belt and braces.

Claude's plan wasn't what I expected. It doesn't watch video the way I do. It works from the audio, turned into a transcript, plus still frames it can pull when it needs to. For a podcast, that's most of what matters. The risky stuff is in what people say.

The heavy lifting happens on my laptop, through a free tool called ffmpeg. When Claude told me to install it, my first reaction was "I don't want to install some weird-ass malware." Its answer was reasonable: ffmpeg has been around since 2000, apps like VLC and OBS are built on its libraries, and it installs with one command through winget, Microsoft's package manager. I didn't just take its word for it, though. I googled it, read up myself, and got comfortable. That's become my rule with AI: trust it, but check the bits that matter.

Whisper, Running on My Own Laptop

Step one was pulling the audio off the SD card. The first attempt made a full preview video and crawled along at half real-time, because decoding 4K is heavy. Claude killed it and switched to copying the audio track only: about 13 minutes, mostly waiting on the SD card.

For the transcript, it turned out the ffmpeg build I'd installed includes Whisper support. Recent ffmpeg versions can run OpenAI's Whisper model locally (via whisper.cpp), if the build includes it. So the original audio and video never left my laptop. To be clear about what did: the transcript went to Claude to work with, plus a few still frames for designing the overlays and thumbnail.

The model comes from Hugging Face. I used it a bit about three years ago when I was at TRACK, and the easiest way to describe it is GitHub for AI models. Reading the script Claude wrote, the whole thing was: download one model file in the right format, point ffmpeg at it, done. Seeing how easy it is to pick up someone else's model and use it opened up a lot of ideas.

Claude suggested which model to grab. I pushed back: how do you know that's the best one, or are you guessing from old training data? It admitted it had picked from memory, then checked the actual model list and some benchmarks, and changed its pick to large-v3-turbo: a similar download size, faster, and on the benchmarks more accurate. Transcribing 90 minutes took 26 minutes. I made a coffee.

Jargon, Plus Two Accents

Daniel's Indonesian too, and we were throwing product names and tech jargon around the whole time. Whisper did its best. The transcript came back a bit weird.

This is where having a language model read it made the difference. Search that transcript for "VBA" and you'll find nothing. But Claude reads "the macro PBA" in the middle of Daniel talking about Excel and his audit work, and knows it's VBA.

Intent, Not Keywords

I gave Claude a simple brief. Keep the tech conversation. Cut anything that isn't about tech, especially anything about people who aren't part of the conversation. Aim for around 30 minutes. It's going public.

Notice what that brief isn't: a list of keywords. "Remove anything where we name people who aren't relevant to the podcast" is intent, not a search term. You can't Ctrl+F for it. A keyword search would miss half the names, especially with Whisper mangling everything, and flag the wrong ones. But a language model reads the whole conversation for meaning, not just words. It's not flawless, as you'll see, but it's close. It knows the difference between a story about a mutual mate and a chat about a product. I could just say what I wanted, and it understood what I meant.

It came back with the conversation broken into sections, what to keep, what to drop, and a short list of things it thought I should decide myself: names of people who had nothing to do with the topic, a few offhand comments about employers, and a claim about a named firm that didn't need to be in there. I made the calls. Nothing got cut without my sign-off.

It kept the context too. A few of its first trims would have cut someone off mid-thought, or left the next bit not making sense. It spotted those and put them back, even though that made the episode longer. It also got its maths wrong once: its first estimate said my picks came to 33 minutes, and the real number was 57. It told me straight away and gave me options. We landed at about 50 minutes, a normal length for a long-form chat.

Then it put the whole edit into a shareable doc: kept lines in normal text, cut lines struck through, timestamps on everything. I read the edit in ten minutes instead of watching it for ninety.

The worry about jump cuts

My biggest concern was that it would chop out every "um" and "yeah, nah" and leave something robotic. It didn't. It cut whole chunks, ten seconds or longer, and left the natural back and forth alone. Every cut lands in an actual pause in the audio. And because I filmed in 4K and exported in 1080p, it zooms in slightly on every second section, so a cut looks like a camera change instead of a glitch. I can tell it's there. It doesn't bother me.

Let the Graphics Card Do It

Before committing, it rendered a one-minute test either side of a cut. The cut looked fine, and in that test my laptop's NVIDIA card decoded the 4K footage about twice as fast as the CPU. The full render went from "leave it overnight" to about 30 minutes.

The clever bit: it doesn't render one big file. It renders 18 chunks, all with identical settings, then joins them without re-encoding. If I hate one join, only the two chunks either side get redone, so a fix takes about ten minutes. That's what made me relaxed about saying "go". I ran it at low priority and thought, well, I'm waiting anyway, might as well get some work done. Not recommended if you're trying to work less. This stuff gets a bit addictive.

Making It Not Look Like Two Blokes in a Cafe

Once the edit was done, I asked if it could add my logo. That turned into a title banner, name tags for Daniel and me, a chapter card at the start of each section, and a small kepakisan.co.nz watermark. It mocked them up on a real frame, in my site's colours, before rendering anything.

I changed my mind on the first banner. It said "two ex-PwC Indonesians", but Daniel is back at PwC, and leading with "Indonesians" might make people expect a video in Indonesian. Two minutes later it said "two old PwC mates having a yarn", true for both of us. Then it cut three short clips for LinkedIn with hand-cleaned captions, wrote the YouTube description and chapters, and made the thumbnail.

Honestly, the bit that amazed me most was the chapter cards. I never told it where each section started. It worked that out from the transcript and put each card right where the conversation changed topic. That means it knows what's being said and when. So in theory it could pop up an Azure or Power BI icon the moment one of us says the word. On this podcast that would be overkill, but you get the idea. For a monologue about a bit of tech, with icons and a diagram building as I explain it, I don't see why not. I haven't tried that yet.

What It Didn't Do

It doesn't watch the video, either. It can check frames when you ask, but it isn't scanning every second for a phone notification on screen. That visual check is still on me. So are the calls that matter most: what's private, what's fair to publish, whether something said is actually true. That's not polish, that's responsibility, and it stays with me.

Could I make it perfect? Sure. But that last 5% is taste, and taste is my job. Good enough plus my evenings back is a trade I'll take every time. Claude did the first pass, and the first pass is the part that used to eat my weekend.

What This Means at Work

That was the point of the test. Now I know how good it is, I can see where it fits at work.

Most organisations make far more video than they publish: training, inductions, safety briefings, leadership updates, recorded workshops, and internal promo like introducing a new system to staff. Most of it is watched internally and needs to be clear, not cinematic. Right now it either queues up for whoever does video, someone loses an afternoon in an editor, or it never gets made.

One note: screen-based how-tos, like walking through an app, are better done the screenshot way I mentioned at the start. This workflow is for recorded footage of people.

For internal content, good enough visually is genuinely good enough: captions, a clean cut, a title card, chapters. It still has to be accurate, and someone still has to review it. Captions alone are worth it, since people watch on mute and plenty of staff prefer reading along.

For external content, I'd still want it done properly. That's where a marketing person, or a professional, earns their keep. To be frank, I don't see this replacing them. Shooting it right on the day, the angles, the taste to make it feel like a story: that's the craft, and bad footage edited by a script is still bad footage. What changes is that they stop being the bottleneck for every internal video.

Mostly, though, it means I'll actually keep doing this. Have a coffee with a mate, press record, let the editing take a few hours instead of a weekend. So expect more videos from me, probably with fancier graphics, maybe starting with how I make those how-to videos. The idea is mine, the voice is mine, the face is mine. The background might not be.

And if any of this sparked an idea, or you just fancy a yarn, flick me a message through my contact form. Fair warning: you might end up on the podcast.

Watch It

Here it is: Daniel and me on Fabric vs Databricks, who should really own cloud costs, AI agents poking around in your Azure, and what all this means for juniors coming into the industry. Fair warning, there's some swearing. It's two mates talking.

Video not loading? Watch it on YouTube.

Cheers to Daniel for the yarn. Views are our own, not our employers'.