WEBVTT 00:00:00.000 --> 00:00:03.538 Hey. So somebody asked me the other day... can you actually make a podcast 00:00:03.538 --> 00:00:07.076 with AI? Like, the whole thing? And here's the thing — you're listening to 00:00:07.076 --> 00:00:10.759 one right now. This whole show is made with AI. So... yeah. The answer's yes. 00:00:10.759 --> 00:00:13.328 But there's a whole spectrum of what "yes" looks like. 00:00:13.328 --> 00:00:17.602 This is AI Honestly. I'm Drew. Each episode I take one real question from a 00:00:17.602 --> 00:00:22.106 founder or business owner — and tell you what's real, what's not, and what it'd 00:00:22.106 --> 00:00:24.474 actually take to build. No hype. No DOOM. 00:00:29.474 --> 00:00:29.474 00:00:29.474 --> 00:00:33.305 So I get this question because making a podcast sounds like a lot of work. 00:00:33.305 --> 00:00:37.293 Recording is hard. Editing is hard. Writing a script is hard. And most people 00:00:37.293 --> 00:00:41.438 who have something to say don't have a free afternoon every week to sit down and 00:00:41.438 --> 00:00:45.111 do all that. So when they hear "AI can make your podcast," what they're 00:00:45.111 --> 00:00:48.995 wondering is... can I skip the tedious part and still end up with something 00:00:48.995 --> 00:00:49.257 good? 00:00:49.257 --> 00:00:53.251 And the answer is yes. With a catch. Let me walk you through how this show 00:00:53.251 --> 00:00:57.245 actually gets made, because it's a pretty good example of the whole range. 00:00:57.245 --> 00:01:00.972 Here's the flow - I talk. That's it, that's my part. I open the voice memo app 00:01:00.972 --> 00:01:04.894 on my phone and I just... ramble. I could be in traffic with sirens going. I could 00:01:04.894 --> 00:01:08.524 be in the car, I could be at my desk. Doesn't matter. I go back and forth, I 00:01:08.524 --> 00:01:12.349 repeat myself, I lose my train of thought — none of it matters, because the next 00:01:12.349 --> 00:01:13.220 steps clean it up. 00:01:13.220 --> 00:01:17.664 The voice memo app gives me an automatic transcript. I take that transcript and I 00:01:17.664 --> 00:01:22.163 feed it into AI, and that's where it gets turned into a script. Now, I've set up a 00:01:22.163 --> 00:01:26.552 bunch of guidelines ahead of time. What an episode is, what it should cover, the 00:01:26.552 --> 00:01:30.162 tone, all that. So it's not just summarizing me. It's reshaping my 00:01:30.162 --> 00:01:33.051 rambling into something that actually holds together. 00:01:33.051 --> 00:01:37.302 But that script sounds like AI if you don't push it the extra mile. You know 00:01:37.302 --> 00:01:41.666 the tells. The em-dash with the little descriptor after it. The line that goes 00:01:41.666 --> 00:01:46.087 "it's not this - it's that." All that stuff. People can smell it. So I've got a 00:01:46.087 --> 00:01:50.508 long list of things I tell it not to do, and that gets me maybe 80% of the way. 00:01:50.508 --> 00:01:54.702 The last 20% is me, by hand, just chopping. Cutting the AI-scent stuff out. 00:01:54.702 --> 00:01:58.840 And that part actually works for me, because I'd rather a show be a little 00:01:58.840 --> 00:02:01.844 abrupt than a little eloquent. I don't need the fluff. 00:02:01.844 --> 00:02:05.994 Then the script goes into ElevenLabs. That's the voice part. It's a clone of my 00:02:05.994 --> 00:02:10.144 voice, and it reads the script back to me. And ElevenLabs is also where I do my 00:02:10.144 --> 00:02:14.241 editing — I'm listening live, chopping things, rearranging, and that's where I 00:02:14.241 --> 00:02:18.445 drop in my theme music too, which is a little AI-generated track. I just tuck it 00:02:18.445 --> 00:02:20.094 under the front of the episode. 00:02:20.094 --> 00:02:24.289 Here's a thing about voice cloning, though. You'd think you'd want the most 00:02:24.289 --> 00:02:28.598 polished, professional clone possible. You can do that. But a really polished 00:02:28.598 --> 00:02:33.077 clone often has this... AI scent to it. Sounds too clean. In my case?? My actual 00:02:33.077 --> 00:02:37.669 delivery is kind of monotone and flat. So when I run it through AI, the quirks get 00:02:37.669 --> 00:02:42.261 masked, because flat is just how I talk. It works in my favor. That's not gonna be 00:02:42.261 --> 00:02:46.570 true for everybody — but it's worth knowing that "more polished" isn't always 00:02:46.570 --> 00:02:47.307 "more real.". 00:02:47.307 --> 00:02:49.815 Okay. What it'd take if you wanted to do this yourself. 00:02:49.815 --> 00:02:53.680 The voice part — ElevenLabs — is the one real cost. You're looking at their 00:02:53.680 --> 00:02:57.806 Creator plan, which is twenty-two dollars a month. That's the tier that gets you 00:02:57.806 --> 00:03:01.933 the good voice cloning. And it runs on credits. Here's the thing to watch: every 00:03:01.933 --> 00:03:05.902 time you edit and re-render the voice, you burn credits. So if you're doing a 00:03:05.902 --> 00:03:09.872 weekly show with a lot of fiddly editing, you can chew through your allotment 00:03:09.872 --> 00:03:13.685 faster than you'd think, and then you're paying overages. Budget for that. 00:03:13.685 --> 00:03:17.808 The rest of my pipeline is basically free. After ElevenLabs, I export two 00:03:17.808 --> 00:03:22.102 files — the MP3 of the episode, and a VTT file, which is the transcript with 00:03:22.102 --> 00:03:26.511 timestamps. I hand those to my pipeline in Claude Code and I just say, publish 00:03:26.511 --> 00:03:30.520 this episode. And it does the whole thing. It writes the show notes. It 00:03:30.520 --> 00:03:35.158 builds the new entry in the feed, the one Apple and Spotify and Amazon pick up. It 00:03:35.158 --> 00:03:39.452 creates a page on my website for the episode — which is partly SEO, but it's 00:03:39.452 --> 00:03:43.919 more than that, people actually go back and read these. And it publishes all of 00:03:43.919 --> 00:03:44.892 it to Cloudflare. 00:03:44.892 --> 00:03:48.668 And Cloudflare makes this nearly free. Normally, hosting a podcast — unless 00:03:48.668 --> 00:03:52.700 you're locked into Spotify's ecosystem — runs you at least twenty bucks a month. 00:03:52.700 --> 00:03:56.680 Cloudflare is so cheap that storing this stuff is basically pennies. Maybe if I 00:03:56.680 --> 00:04:00.661 hit a hundred episodes and I'm storing all of it, I'd start paying something. I 00:04:00.661 --> 00:04:02.447 don't know. But right now? Pennies. 00:04:02.447 --> 00:04:06.743 There's even some simple analytics baked in. A little dashboard — filters out 00:04:06.743 --> 00:04:10.643 bots, deduplicates, shows me who downloaded an episode, who viewed the 00:04:10.643 --> 00:04:14.656 page. Nothing fancy, but it's there. And the system also pulls out clips 00:04:14.656 --> 00:04:19.008 automatically — finds the interesting soundbites, drops them on a graphic with 00:04:19.008 --> 00:04:23.417 subtitles and the episode question up top. So I've got social posts ready to go 00:04:23.417 --> 00:04:24.830 without doing extra work. 00:04:24.830 --> 00:04:27.616 Now — what to watch for, beyond the credits thing. 00:04:27.616 --> 00:04:32.106 Building the pipeline itself takes real time up front. The per-episode workflow 00:04:32.106 --> 00:04:36.539 is fast — ten minutes to talk into my phone, then up to a half hour of editing 00:04:36.539 --> 00:04:41.030 before it's live. But the pipeline that makes that possible? That was the work. 00:04:41.030 --> 00:04:45.635 So if you're picturing zero effort, it's zero effort after you've done the setup. 00:04:45.635 --> 00:04:50.104 The other catch. The hard part of a podcast was never the production. And AI 00:04:50.104 --> 00:04:54.871 just took the production part and made it accessible. What it didn't solve — what 00:04:54.871 --> 00:04:59.578 nothing solves — is the marketing. The distribution. Getting anybody to actually 00:04:59.578 --> 00:05:03.868 listen. That's still the real job. The making is largely handled now. The 00:05:03.868 --> 00:05:05.000 being-heard is not. 00:05:05.000 --> 00:05:09.288 So... can you make a podcast with AI? Yes. And it's a genuinely good setup, 00:05:09.288 --> 00:05:13.865 especially compared to cloning yourself on video, which is way harder — voice is 00:05:13.865 --> 00:05:18.269 much easier than video. If you've got something to say and you don't have the 00:05:18.269 --> 00:05:22.789 time to run a weekly show the old way, you can replace most of that with AI and 00:05:22.789 --> 00:05:26.787 end up with something real. This is actually episode two, so I'm still 00:05:26.787 --> 00:05:30.495 finding out how it holds up. But the production? That part works. 00:05:30.495 --> 00:05:34.865 That's it for this episode. If you've got a question you've been wondering about... 00:05:34.865 --> 00:05:39.129 can AI do this thing?? send it to me. The address is in the show notes. I'm Drew. 00:05:39.129 --> 00:05:40.248 This was AI Honestly.