Vibecoding 003 — Extracting One Usable Idea From Every Podcast Episode
Table of Contents
One usable idea from every episode #
Project #3 of the 100 Vibe Coding Projects challenge
Think Fast, Talk Smart is Matt Abrahams’ podcast on communication — 270+ episodes, each guest offering their best advice on how to speak, listen, and connect more effectively. The problem: that advice is scattered across hundreds of hours of audio. Nobody’s reading the back catalogue.
The prompt was: extract the core advice from each episode, compress it to 7 words or less, and put it on a screen.
The result is a 6-minute video montage covering 165 episodes — each card showing the episode number, topic category, guest name, and their advice in large bold type, colour-coded by theme (teal for Communication, red for Conflict, purple for Mindset, gold for Feedback).
Version 2 #
The first version worked but felt clinical. Version 2 rebuilt the music from scratch: warmer, more orchestral, closer to a podcast about human connection than to a product demo. The backgrounds became illustrative rather than blank: each category got its own abstract image (sound waves for listening episodes, mountain silhouettes for leadership, a neural-network pattern for mindset). Transitions slowed down and softened.
The honest verdict: technically better, emotionally closer. Still not perfect — a human voice reading the advice would transform it from a slideshow into something worth watching twice.
Takeaways #
Narrative needs to stitch together all of this advice. A string of bullet points across 165 slides is data, not a story. The raw material is good; the frame around it still needs work.
The security angle #
Extracting structured insight from unstructured text across 165 podcast episodes is the same pipeline used in threat intelligence processing: take raw text (blog posts, threat reports, OSINT feeds) and pull specific signals out of it. The difference is the signal type — guest advice instead of indicators of compromise.
The technique is identical. If you can prompt an LLM to extract “the advice each guest shares in 7 words or less,” you can prompt it to extract “the malware family, affected platforms, and recommended mitigations” from a threat report. Same pattern, different domain. This project was a useful reminder that the automation skills transfer directly.