Dheeraj Sharma and I just wrapped season one of One Shot Show with episode 9, and we ended on a guest I have wanted to dig into for months: Michael Simmons.
Michael writes for Forbes, Harvard Business Review, and Fortune. His articles have been seen over 100 million times. On Medium, his average post pulled more than 250,000 views. For the past decades, he has been obsessed with one question: how do you build a system that reliably produces blockbuster content, not just one good article?
These days he writes Blockbuster Blueprint, where he breaks down how to apply that system using AI. He is also building Cozora (where I’m an active creator and contributor), an AI community where members learn directly from the creators behind some of the most interesting AI workflows on the internet. Michael co-hosts weekly masterclasses there with practitioners walking through how they actually build.
A few weeks ago, Michael invited me on his livestream to share how I run AI Maker inside Claude Code. We talked about skills, MCP, the CLI, and how I have been wiring most of my newsletter workflow into one harness. After that conversation, one thing was clear: we both love Claude Code, and we both keep finding new ways to bend it around our own work. In case you missed it, you can watch the session here:
So this time I invited him on. Because in that same conversation, Michael casually mentioned a workflow that I could not stop thinking about. He uses podcasts as a research source for his Claude Code setup. Not as background entertainment. As actual raw material that flows into the same vault his agent reads from when he is drafting a Forbes piece or a newsletter post.
That is what I wanted him to walk through on this episode. And the answer he kept coming back to was podcasts.
Podcasts Are a New Source for Your LLM Wiki
When Andrej Karpathy posted his LLM wiki idea, most of us latched onto the same intake sources. Articles. Web clippings. Book highlights. Old documents. Things you read on the internet, saved to Obsidian, and pointed Claude Code at.
I wrote about it in this post:
That works. I do a version of it. I’m sure a lot of you do too.
But there is a whole other channel of information most of us already consume every day that almost never makes it into the vault: podcasts.
Think about how much of your AI learning right now happens through audio. If you’re like me, you might be listening to Lex Fridman, Lenny’s Podcast, Dwarkesh, Moonshot by Peter Diamandis, etc. The point is: podcasts are where the most interesting people in AI think out loud, often in more depth than they ever do in writing. You listen on a walk, on a commute, while cooking. You hear something that lands. An hour later, you cannot remember which episode it was, let alone the timestamp.
That insight is gone. It never enters your second brain. It never becomes raw material for Claude Code to draft from.
Michael’s setup fixes that. He treats podcasts as a first-class source for the LLM wiki, on the same level as articles and books. Every episode he listens to is a potential stream of clips, transcripts, and atomic notes flowing into the same vault his Claude Code agent reads from.
That single shift, treating audio as a real intake source instead of background entertainment, is what I want to walk through.
💡 A quick related note…
I found this entire Claude free guide library. It maps out everything you need to know about Claude. From beginner to expert. 10+ free guides. Claude 101, Cowork, Skills, Code, Teams.
Access it here → claude101.com (there is no catch, no paywall).
How the Podcast Pipeline Actually Works
The tool doing the heavy lifting is Snipd. I had heard of it. I had not really tried it. After this conversation, I am setting it up this weekend.
Snipd is a podcast app, but the part that matters is what happens when you tap the snip button. Snipd uses AI to figure out the actual context: where the idea started, where it ended, who said it. You get a clean clip with a full transcript already attached, and a star button to mark the ones you want to keep.
That is the unlock. Every podcast you already listen to becomes a stream of small, time-stamped, transcribed, attributed notes. Instead of typing and summarizing manually, you just keep listening as usual and tap when something resonates.
Then Snipd syncs the starred clips into Obsidian, Notion, or Readwise. A few clicks to set up. After that, every clip you star while walking the dog or commuting shows up in your vault, ready for Claude Code to read.
Michael’s vault now has around 11,000 notes in it. A lot of them came in this way.
A few details from the livestream demo that I had not seen before:
You can follow specific guests, not just shows. Michael follows Sam Altman, Dario Amodei, Demis Hassabis, Andrej Karpathy, and a few hundred others. When any of them shows up on a podcast he does not normally listen to, the episode lands in his queue.
You can chat with an episode. Remember a guest said something about agent harnesses but cannot find where? Ask the episode. It jumps you to the timestamp.
You can upload outside content too. Books from Libro.fm as MP3. YouTube videos. Even long-form articles. Michael built a small Claude Code skill that takes an article he wants to read, sends it to ElevenLabs, gets the MP3 back, and pushes it to a personal RSS feed that shows up in his podcast player. The article becomes a podcast he can clip from.
Once the clip is in the vault, Claude Code can do anything with it. Michael uses the same vault to draft articles, share clips into a paid subscriber WhatsApp group, and run a weekly summary that stitches starred clips together.
The system is doing the work. He is just listening.
Why Podcast Clips Are a Better Than Book Highlights and Saved Tweet
I want to call out one thing that I think gets missed when people talk about second brains.
Most second brain advice treats all sources the same. A book highlight, a tweet, a clipped article, a meeting note. Throw them all in, link them up, query later.
In practice, the source matters a lot for what Claude Code can actually do with it.
A book highlight is just a sentence on its own. No speaker, no context, no surrounding argument. A saved tweet is short and usually missing the thread. A clipped article often loses the part of the argument right before the line you cared about.
A podcast clip is structurally better. It has a speaker you trust. It has a timestamp. It has the surrounding 60 seconds of context. It has a full transcript. And because someone said it out loud, it usually carries the actual argument, not just the conclusion.
When Claude Code searches across 11,000 notes looking for material to draft a piece, that extra context is what lets it pull genuine evidence instead of generic summaries.
This is also why Michael does not just use Snipd for podcasts. He keeps converting other formats into audio so they enter the system through the same pipe. Every input ends up with the same shape: speaker, timestamp, transcript, clip. That consistency is what lets the second brain compound instead of becoming a junk drawer.
How to Add Podcasts to Your Second Brain This Week
If you already have an LLM wiki or second brain running on articles and old documents, the good news is you do not need to redesign it. Instead, you can simply plug podcasts in as one more source.
Here is what I would do this week:
Install Snipd and import the shows you already listen to.
Pick three guests you trust and follow them as people, not just their shows.
For the next week, tap the snip button every time something hits. Star the ones you want to keep.
Set up Obsidian or Notion, then start syncing them together so the clips land in your vault automatically
Then, and only then, point Claude Code at the vault and ask it to surface patterns or draft something from your clips.
That is the loop. The reason it works is that you are not adding a new habit. You are bolting capture onto a habit you already have.
Other Threads From the Conversation
We covered more ground than just Snipd. A few threads worth pulling on, each one probably its own future post:
Codex is now a real alternative to Claude Code. I opened the call telling Michael I have been going crazy with Codex lately. Not the CLI. The app. Claude Code still has the better agent harness and structure in my opinion, but Codex feels like a polished super-app you can actually live in, instead of a terminal. You can even open Claude Code inside Codex and run both models side by side.
How we use AI has been evolving fast, and Claude Code is the latest jump. Think about how this has changed in just a couple of years. First it was pure question and answer. Open ChatGPT, ask something, copy the output into wherever you actually do your work. Then it became AI automation. You opened Make.com or Zapier and built a deterministic workflow by chaining steps together: this trigger fires, then this AI step runs, then the result lands in this app. It worked, but every workflow was a small project. Michael told a story on the call about spending eight hours one day in Make.com trying to build a single workflow. Now we are in a third phase. Claude Code, Codex, and tools like them can do most of those workflows directly, just by being asked. The same thing Michael spent eight hours wiring up in Make.com, Claude Code did in 15 minutes. The reason this matters is not just speed. It is that the cost of trying a workflow has dropped to almost nothing, which means you experiment with way more of them.
AI video is still expensive enough to matter. Dheeraj walked through his HeyGen experiment. Around $5 per minute at scale via the API. A 30-minute video burns through a $30 plan in one shot. Useful if you already have a funnel and a high lifetime value per viewer. Painful if you are starting out and trying to automate your YouTube channel from zero. The “AI does everything for me” pitch in YouTube videos hides this part.
The uncanny valley is closing faster than people think. Michael described three stages of AI content. Stage one is obviously bad and no one engages. Stage two is clearly AI but the value is unique enough that people watch anyway. He used the Epstein Files Podcast as an example, where someone used AI to chew through three million leaked files and shipped 150 fact-based episodes in one drop. Stage three is when AI content gets more personalized, more researched, and easier to update than anything a human can produce one-to-many. We are somewhere in stage two right now.
A Note on the Bigger Pattern
One thing Michael said near the end stuck with me. He thinks the cost of producing high-quality content, once your funnel is set up, is going to keep dropping. The bottleneck for most of us was never the writing; it was the raw material and how we stitch it all together.
That is the part the podcast pipeline really shifts. The audio you are already consuming becomes raw material your Claude Code agent can actually work with. Quotes you would have forgotten turn into clips you can search, link, and draft from.
Articles and web clippings still belong in the wiki or second brain. Podcasts just join them as one more source, and probably the one with the highest ratio of insight to effort once it is set up.
That is the version of the LLM wiki or second brain I am actually going to use.
Episode Details
One Shot Show, Episode 9. Season 1 Finale.
Live every Wednesday at 10:00 AM EST on Substack.
Guest: Michael Simmons, founder of Blockbuster Blueprint. Forbes, HBR, and Fortune contributor.
Timestamps
00:00. What is exciting in AI right now: Codex vs Claude Code
04:00. Welcome and season one wrap
06:22. Michael’s background and the blockbuster framework
09:00. From ChatGPT to Make.com to Claude Code
14:52. Connecting writing to a second brain
17:00. Michael’s news-to-article workflow
18:30. Skills, chained skills, and the second brain
20:00. How to actually start a second brain
28:30. Snipd demo: AI clipping, follow-by-guest, audio hub
36:00. Skills for ElevenLabs, YouTube, and weekly summaries
41:00. Build vs buy and the cost of AI tools
50:00. AI-generated content and the uncanny valley
55:00. One thing to start this week
58:00. Codex and Claude Code as a bridge
Resources Mentioned
Snipd (snipd.com). AI podcast app with smart clipping, follow-by-guest, chat with episode, and Obsidian/Notion sync. Upload your own audio at upload.snipped.com.
Andrej Karpathy’s LLM Wiki. The second brain idea Michael built around.
Libro.fm. Audiobook MP3 source you can upload to Snipd.
ElevenLabs. Text-to-speech, used in Michael’s article-to-podcast skill.
HeyGen. AI avatar video tool Michael uses for weekly summary clips. Around $30 per month for 10 minutes of video, or roughly $5 per minute via API.
Obsidian / Notion. Second brain vaults, both sync directly with Snipd.
Claude Code. Michael’s current agent harness for skills, search, and drafting. Opus 4.6 was the unlock for him.
Codex. OpenAI’s coding agent app. Came up as a complement to Claude Code, especially around weekly limits and bridging plans.
Blockbusters by Anita Elberse. The book that shaped Michael’s thinking about high-quality content as a strategy.
