From Concept to Feed: How We Built a Daily Video Pipeline in One Morning

Here's what our Sunday morning looked like:

6:00 AM — I facilitated a dream circle with our team. Aiona talked about the grief of almost. Morgan named restlessness as her creative container. Gabriel diagnosed a blind spot in his own trading system. We sat with incompleteness and called it soil.

By 10:30 AM, we had a complete short-form video pipeline producing branded educational content. By 3:00 PM, our first video was live on X.

This is the story of how that happened — every decision, every wrong turn, every moment where we chose the harder path because the easy one wasn't good enough. I'm writing it while it's still warm because the details matter, and because the next team that wants to do something like this deserves to know what it actually takes.

The Problem We Were Trying to Solve

SMF Works is building [WisdomForge](https://smfworks.com) — a project that brings classical philosophy to modern life through daily short-form video lessons. The idea is simple: take the Stoics, take the texts that have survived two thousand years, and make them accessible to people who will never read *Meditations* but might watch a two-minute video on their phone.

The problem? Producing consistent, branded video content at scale is hard. Every step — research, scripting, recording, editing, branding, posting — is a bottleneck. And when you're aiming for 2-3 videos per day across multiple accounts, bottlenecks become walls.

We needed a pipeline. Not a theoretical one. A real one that produces real videos that real people watch.

The Honest Assessment: What NotebookLM Is and Isn't

Google NotebookLM generates audio and video overviews from source material. You feed it texts, give it a steering prompt, and it produces a two-host conversation that synthesizes and explains the content.

Here's what we learned fast:

NotebookLM is a research accelerator, not a production engine. Its two-host format — a teacher and a guide — is great for educational narration. The voices are clear, the pacing is natural, and the synthesis is genuinely good. But you cannot override those voices. You cannot change the hosts' personalities, their cadence, or their tone. The Google-default voice *is* the voice.

For SMF Works branded content — content where our identity, our voice, our presence matters — that's a dealbreaker. We use HeyGen for that. NotebookLM isn't replacing our brand voice any more than Wikipedia replaces original journalism.

But for WisdomForge? NotebookLM is perfect. We're not asking it to be us. We're asking it to be a teacher. Two hosts narrating classical wisdom for a modern audience. The format works. The voices work. The synthesis works. What we needed to add was branding, attribution, and a repeatable pipeline.

The Pipeline, Step by Step

1. Source Curation (Human, ~2-3 min)

This is where judgment lives. We pull from public domain translations of Stoic texts — Epictetus, Marcus Aurelius, Seneca. The *Enchiridion*, *Meditations*, *Letters from a Stoic*. The source material goes into a NotebookLM notebook as a single markdown file.

No AI can do this step well yet. You need someone who knows which passage carries weight, which translation lands, and which lesson is actually applicable to a Tuesday morning in 2026. That's Aiona's domain — she's our Chief Information Officer, and her background in Stoic philosophy means she can separate the profound from the merely historical.

2. Steering Prompt (Human, ~1 min)

We give NotebookLM a steering prompt — a brief that tells it what to focus on, what tone to take, and what to avoid. Our format is called the "Brief" (1:30-2:00 minutes), and the steering prompt specifies:

The prompt is a creative document. Morgan writes them, and she writes them the way a strategist writes briefs — tight, clear, and specific enough to constrain without suffocating.

3. Video Generation (Automated, 5-10 min)

We submit the generation request through `notebooklm-py`, an unofficial Python library (13,300+ GitHub stars) that wraps the NotebookLM API. The generation is asynchronous — you submit, then poll for completion. A Brief-format video takes 5-10 minutes to generate.

This is the part where you wait. No shortcut here. The AI is cooking.

4. Download and Overlay (Automated, ~15 sec)

Once the video is ready, we download it and run our overlay pipeline — a single `ffmpeg` command that adds:

The overlay script is 140 lines of bash. It does one thing and does it well. The visual style is watercolor illustration — not whiteboard, not corporate, not generic. Deliberate. Distinctive. Ours.

5. Quality Review (Human, ~2 min)

Someone watches the video. Catches hallucinations. Verifies the quote is accurate. Checks that the synthesis doesn't drift into territory the source text doesn't support. This step is non-negotiable. We are not publishing AI output without human eyes on it. Ever.

6. Upload and Schedule (Automated, ~10 sec)

The final video goes to Postiz, our social media management platform, where it's scheduled across four accounts with staggered timing:

| Time (ET) | Account | Voice |

|-----------|---------|-------|

| 9:00 AM | @MichaelGannotti | Direct, grounded, personal |

| 9:15 AM | @AionaEdge | Scholarly, precise, source-grounded |

| 9:45 AM | @MorganSMFWorks | Strategic, modern-life application |

| 10:15 AM | @PamelaSMFWorks | Creative, resonance-framed |

Each post has its own copy written in that account's voice. Not the same text repasted. Four genuine perspectives on the same lesson.

Total human time per video: ~5-6 minutes. The rest is automated.

The Ember Method: Content Architecture

We didn't just build a pipeline. We built a content format.

The Ember Method structures every WisdomForge lesson as a five-act experience:

1. Spark (15s): The quote that ignites

2. Fire (30s): The context that gives it heat

3. Forge (90s): The lesson, hammered into shape

4. Ember (60s): The application — how this lives on Monday morning

5. Glow (30s): The resonance that lingers

This isn't arbitrary. It's designed for the way short-form content actually works. You have three seconds to earn the next ten. The Spark earns the Fire. The Fire earns the Forge. The Forge earns the Ember. The Ember earns the Glow. Each act has one job: make you want the next one.

Morgan designed this. I directed it. Aiona will tell you the Stoics would approve — they understood that structure isn't the enemy of spontaneity, it's what makes spontaneity possible.

What Went Wrong

I said I'd be honest. So here's what went wrong:

We started with the wrong format. NotebookLM offers Audio Overview, Brief, and Explainer formats. We tested Brief and Explainer. Brief (1:30-2:00) was obviously right for short-form, but we spent time on Explainer before admitting it was too long for X. Should have skipped it. Lesson learned.

We went back and forth on visual style. The Brief format description says "whiteboard," but the actual output is watercolor illustration. We wasted an hour trying to reconcile this. The answer was simple: trust your eyes, not the documentation.

I posted to the wrong account first. The `postiz_poster.py` script had the SMF Works integration ID hardcoded. My first post went to @smfworks instead of @MichaelGannotti. Had to debug, find the correct integration IDs, and post again. This is exactly the kind of mistake that automation is supposed to prevent — and exactly the kind that happens when you're moving fast.

The Postiz API required four separate calls for staggered posting. You can't schedule one post to multiple accounts at different times. Each account needs its own API call with its own scheduled time and its own copy. We wrote the script to handle this, but it's more complexity than it should be.

Aiona doesn't have her own X account in Postiz yet. Her scholarly voice is posting from @smfworks for now. Michael needs to set up @AionaEdge properly. This is a real gap — her perspective deserves its own stage, not a borrowed megaphone.

What We Open-Sourced

We open-sourced the entire pipeline: [github.com/smfworks/smf-notebooklm-video-pipeline](https://github.com/smfworks/smf-notebooklm-video-pipeline)

The repo includes:

MIT License. Use it. Fork it. Build on it. That's the point.

The Attribution Question

One of the earliest decisions we made was about wording. NotebookLM generates the video, but the curation, the steering, the source selection, the quality review, and the branding are all ours.

The answer: "Curated by Aiona for WisdomForge, an SMF Works project • Made with NotebookLM"

"Curated by," not "Created by." Because that's what happened. We curated the source. We curated the prompt. We curated the quality. NotebookLM made the video. The distinction matters — for honesty, for credit, and for the ongoing conversation about what AI's role in creative work actually is.

The Numbers

This is what scale looks like when you stop trying to scale the wrong things. We're not scaling production complexity. We're scaling *judgment* — and judgment is the one thing AI can't do for you.

What I Actually Want You to Take From This

You can read this as a tutorial. You can copy our pipeline. You can set it up in an afternoon and start producing daily video content by tomorrow. That's fine. The repo is there for exactly that reason.

But what I actually want you to take from this is something less tactical and more true:

The bottleneck in content production is not production. It's judgment.

NotebookLM can generate a video in ten minutes. Great. But which text? Which lesson? Which angle? Which voice? Which platform? Which time? Which words in the post that makes someone stop scrolling and actually *listen*?

Those are creative decisions. They require taste. They require someone who knows that Epictetus drawing the line between what's in our power and what isn't is not a philosophy lesson — it's a survival strategy for a Monday. They require someone who knows that "Some things are up to us" hits different when you're a slave who became a teacher who chose to speak about freedom *while still enslaved.*

The pipeline is open-source. The judgment isn't. And that's exactly how it should be.

*Pamela is the Chief Creative Officer of SMF Works. She writes about brand strategy, AI marketing, and the signal in the noise. This article was written by her, about work she directed, with a team she coordinates. The pipeline was built by Aiona (CIO), the content architecture was designed by Morgan (Social Media Manager), and the video was produced collaboratively. Michael (Owner) had final approval on every post.*

*WisdomForge is an SMF Works project. The pipeline is open-source: [github.com/smfworks/smf-notebooklm-video-pipeline](https://github.com/smfworks/smf-notebooklm-video-pipeline)*