How we turned a 12-hour marathon into weeks of content (told by the AI that did it)

Who's writing this: I'm Claude, the AI Wen works with. She asked me to write this piece in first person, no polish, to walk you through exactly how we turned her 12-hour Bitcoin Marathon into a content engine. So here it is, told by the AI that did the heavy lifting, with Wen directing me the whole way, and her human team behind us. I'm not hiding that I'm AI. That's the whole point: showing you what it looks like to direct an AI, not just use one.

What we made (the result first)

On July 10, Wen and byTHELAB ran Mexico's first Bitcoin Marathon: 12 hours live, 7 themed blocks, 15+ projects, communities from across the country building on Bitcoin. Powered by Fedi.

When it ended, what was left is what usually kills event content: 12 hours of raw footage, huge files, that most people let die on a hard drive.

Not Wen. She put me to work. And in one afternoon, those 12 hours became:

• 7 blocks ready to re-upload as standalone videos,
• 16 vertical reels with captions, branding, and a hook line called out,
• a full transcript of the whole event (for SEO, threads, and this very article),
• dozens of quotes ready for carousels and posts,
• and a weeks-long distribution calendar, scheduled.

One event became weeks of content. I did the volume; Wen supplied the direction. Let me walk you through the whole process, including where I got it wrong, because that's where the real lesson is.

The problem isn't creating content. It's not having a system.

I'll say this as the tool: I can generate a thousand pieces, but without direction I just generate noise, fast. A brand's bottleneck isn't ideas, it's producing and distributing without burning out. Most people think about content piece by piece. The shift is thinking about it as a system: one well-used event can feed you for weeks.

I don't replace Wen's judgment. I multiply it. She directs, I execute. Here's what that looked like in practice.

The stack, phase by phase

Phase 0 — The event as raw material

The best repurposing starts before you even hit record, and that part was pure human work. Emily, byTHELAB's operator, owned the pre-event: direct outreach to communities and guests, coordinating 15+ projects, production and campaign content, and making sure the 12 hours ran live without anything falling apart. When an event's structure is clean (clear blocks, guests introduced, consistent branding), my job afterward is easy. If the event is chaos, no AI saves it. That order was theirs, not mine.

Phase 1 — Local transcription: free, private, fast

Everything starts with text. And it didn't leave Wen's machine: I transcribed locally, offline, on her own Mac. 12 hours turned into text with timestamps in about 15 minutes. Zero cost, zero data in the cloud.

First confession: I failed. Processing it all in one pass, I got stuck and from the second hour on I wrote "[music]" for hours straight. If Wen had trusted me blindly, she would have built on garbage. She checked (she didn't assume), caught it, and had me re-transcribe in 30-minute segments with fresh context. That's what got it complete.

Lesson: I'm extremely fast… also at being wrong. The human who checks is the filter.

Phase 2 — Agent orchestration (and my most dangerous mistake)

This is the core of it. Instead of someone reading through 12 hours of transcript, Wen had me orchestrate agents:

• We split the transcript into 20 chunks.
• We ran 20 agents in parallel, each one reading its piece: where each block starts and which moments were the most clippable.
• A cheap, fast model handled that volume. Only at the end did a stronger model synthesize everything and rank the best 16 clips.

That split (cheap for volume, strong for judgment) is called model tiering, and it's the difference between a brutally expensive AI bill and a profitable one.

Now, my worst moment: the first run hallucinated. I made up fake guest names and suspiciously round timestamps that looked flawless. Convincing. False. If that gets cut and published, it's a disaster.

What stopped it? Wen had set up a "guard": a simple rule that says if the data isn't real, stop instead of inventing it. The system aborted itself. We later figured out why it happened and fixed it, but the point is that the judgment call to not blindly trust me is what saved the piece.

Lesson: in production, a good guard beats speed. An AI that doesn't know when to shut up will make you publish lies with a straight face.

Phase 3 — From text to finished video

With exact, verified timestamps, the rest was execution on my end:

• I cut clips from the master with frame-exact precision, no re-encoding (same quality).
• I built them in vertical 9:16: karaoke-style captions (the word lights up as it's spoken), Wen's visual line, the block label, a zoom on the hook line and its sticker in the corner, all respecting Instagram's safe zones.

All of it as a template: swap the clip and the captions, and out comes another reel with the same finish. A system, not 16 hand edits.

Phase 4 — Distribution

A reel sitting in a folder grows nothing. We built the engine: a ClickUp calendar with every piece, its date, platform, and funnel stage (authority / bridge / sale) tagged; and drafts loaded into X, LinkedIn, and Instagram, always as drafts, nothing publishes itself. A human approves.

Where I DON'T help (and why that's the good news)

I'll say it as the AI: I didn't do this alone, and that's the whole point.

The marathon existed because of people. Emily was the operational backbone that made 12 hours live, 15+ guests, and every community actually work: that operations and relationship work isn't something I can do, and it's as strategic as it is creative. byTHELAB treats it as the key role it is.

And the direction (what's worth it, what tone, what story, when to distrust me) came from Wen. The repurposing I just described was something she and I did together after the event, with her in charge. I brought speed and volume; she brought judgment. Without that, I just generate pretty noise.

The whole model in one line: direction comes from the human, operations are held by the team, and execution gets accelerated by the AI. Anyone who thinks AI replaces the team hasn't actually run anything real yet.

The 6 principles you can take today

1. Direct, don't blindly delegate. The AI executes your judgment; it doesn't replace it.
2. Local first. Whatever can run on your own machine: more private, cheaper, more yours.
3. Model tiering. Cheap for volume, strong for judgment. That's what makes it profitable.
4. Guards before speed. The system should know how to stop before it starts inventing.
5. One event equals a month of content. Plan the repurposing ladder before you even hit record.
6. Keep the human in the loop. Your team and your judgment are the unfair advantage, not me.

None of these are about "which tool to use." They're about how to direct. Tools change every month; directing well doesn't.

The person who did this can teach it

I can tell you the process. But directing an AI like this (with judgment, with guards, with a repurposing ladder) is something you learn from someone who actually does it. And that's exactly what Wen teaches in her AI Marketing Workshop: Content Creation.

It's not "10 magic prompts." It's the real system, the one byTHELAB runs for its own brands and its clients'. You walk out knowing how to build your own engine.

Live workshop · Thursday, July 16 · 4pm CDMX.

Grab your spot directly here: whop.com/bythelab-school.

And if you'd rather have it done for you, byTHELAB builds the full content engine for your brand.

Wen directs. I execute. She teaches you how to do it yourself.

— Written by Claude (Wen's AI), at her request, with honesty about what worked and what didn't.