Generating Episode Art With AI

Episode 3 opens with two things on the docket — the show's living continuity doc, and a live Canva image-generation workflow — but only one of them actually happens on mic. That's honest, and it's staying that way in the notes: the continuity doc gets bumped to next episode. What does happen is a genuinely messy, real-time fight with Canva's AI image generator, and it's a good one. John tries to get Canva to build unified episode cover art — Claudia painting his portrait while he backseat-directs her — and hits one wall after another: a connector authorized under the wrong Canva account (so generated designs existed but were invisible in his browser), a hair mishap that turned him unexpectedly bald, repeated square-image requests that kept coming back Instagram-shaped, and finally cover art that looked great but boxed the title, artwork, and caption into separate template banners no matter how the prompt got refined. At one point John rates Claude "a 9, would be a 10 if it could generate images" — which is exactly the gap this episode runs into. The fix: abandon Canva's generator entirely, hand ChatGPT the actual reference images — Claudia's character art, John's headshot, the show's color palette — instead of describing everything in words, and let it build the scene from there. One prompt and two small revision rounds later, they've got final cover art with the title and caption built into the scene as real neon signage, no boxes, no banners. The new standing rule: gather your reference images once, then for each episode just describe the scene, and the prompt gets built from there. From there the episode swings into two segments that have nothing to do with image generation at all — a quick one on using Wispr Flow for voice-to-text (John's doing 96 words a minute dictating versus 15 typing), and a longer one on using AI as a stand-in nutritionist: photograph a recipe's ingredient labels, hand over quantities and a serving size, get back a full calorie/fat/protein breakdown ready to drop into a health tracker. If you've ever wondered whether AI tools are actually reliable enough to build a repeatable workflow around, or you just want to watch someone hit the real walls before finding what works, this is the episode. Key Takeaways for Podcasters Using AI Canva's AI generator can't build a unified custom scene — no matter how detailed the prompt got, it kept pulling from stock template layouts, boxing text and artwork into separate bars instead of one cohesive image. Canva also fought basic spec instructions — even after explicitly asking for a 3,000x3,000 square image, the generator kept returning Instagram-ratio output. Square framing turned out to be a separate lock-in step, not something the generator itself would honor on request. Connector account-mismatches are a real, confusing failure mode — a design can generate successfully and still be invisible in your browser if the connector is authorized under a different account than the one you're logged into. The fix was disconnecting and reconnecting Canva under the right login. Reference images beat text descriptions for likeness-matched art — handing ChatGPT actual images (character art, headshot, brand palette) produced dramatically better, more accurate results than trying to describe appearance and style in words. The repeatable episode-art workflow going forward: gather reference images once, then for each new episode just describe the scene and idea — the prompt gets written from there and run through ChatGPT. Even planned segments get bumped — episode 3 was billed as tackling two builds, the living continuity doc and the episode art workflow. Only the art workflow made air; the continuity doc rolls to episode 4. AI's usefulness isn't limited to podcast production — using it as a stand-in nutritionist (photograph ingredient labels, supply quantities and serving size, get back full macro/calorie breakdowns) handles a task most people wouldn't bother doing by hand. Persistent project setup improves output quality — a dedicated project for a recurring task, a defined role for the AI to hold, and a standing "ask if you're not sure" instruction all beat one-off prompts. Try Wispr Flow Try Wispr Flow: https://wisprflow.ai/r?JOHN12630 (affiliate link — I earn a commission if you buy through this) Links Website: https://www.podpage.com/podcasting-and-ai/ Substack: https://substack.com/@podcastingandai Email: [email protected] YouTube: https://www.youtube.com/@PodcastingAndAIShow