making content got cheap. distribution didn't.

The brands that win won't be the ones making the most videos. They'll be the ones with a system that knows which ones to bet on.

At 10 accounts, I could still see everything.

Three pieces of content per account, every day. I was reviewing every script, every hook, every frame before it went out. My taste was in the loop. I knew exactly what was going live and why.

Then we crossed 10 accounts, and something broke – not the system, me. I was doing 30+ pieces of content a day, and my review quality was deteriorating. I was catching less. Approving things I wouldn't have approved two weeks earlier. And I wasn't doing anything else – no product work, no strategy, no hiring. Just content review, done badly.

That was the moment I understood the difference between being in the loop and being on it.

I – AI made the cheap part free. The expensive part didn't move.

There's a pattern playing out right now across every industry AI touches, and it's the same pattern every time.

AI compresses whatever was already the scalable part – writing code, generating images, producing video – down toward zero. The thing it doesn't touch is the expensive part: judgment, taste, understanding, knowing what to do with what you made.

Engineering teams in 2026 are living this. The data coming out of teams running AI coding agents shows roughly 4x the raw code output. About 10% more actually shipped value. The code didn't slow down. The humans reviewing it did. Not because they got lazier – because the volume outran their capacity to understand what was coming out. The bottleneck moved downstream to the one thing that never got faster: a person being confident the machine was right.

That's not a coding story. It's a law.

The constraint doesn't disappear when you automate. It relocates.

Content is going through the same thing right now. A skilled operator with the right setup can push 500+ pieces of content a day for a single brand. AI avatars, slideshows, creator-style accounts – all running without manual intervention. Effective CPMs on TikTok are sitting around $0.34, roughly 20x cheaper than Meta. The volume is real. The economics are real.

But here's what I know from building this system: making content was never the expensive part. It just hadn't gotten cheap yet.

II – I wasted 20 million views learning this.

When we were building Flamme, we went hard on memes. They were getting views – 300,000+ on individual videos, millions across accounts. The numbers looked like we were winning.

We weren't. We were just making noise.

Memes convert at roughly 0.05-0.2%. By the time we understood this, we'd burned through the first 20-30 million views without meaningful conversion to show for it. The virality was real. The users weren't coming.

The conversion data by format is brutal when you look at it directly:

• Memes and slideshows: 0.05-0.2%
• Hook + reaction + app: 0.3-0.5%
• Direct product demo: 0.5-1%

The cheapest formats to produce at scale are the worst converters. The math on "I can push 500 slideshows a day at $0.34 CPM" and "I can grow" are not the same equation. One is the cheap part. The other is the part that still costs everything.

We had to learn this by losing the time and the views. Nobody told us – or rather, we weren't asking the right question. We were measuring the wrong output.

III – the VSC framework is how you decide which 550 to bet on.

After the meme period, I built a framework for content selection that I now apply to everything we put out. It has three filters.

V – Viral. The question isn't "can this go viral." The question is timing. There's a curve: signal, climb, peak, plateau, decline. You have to pick the format when it's climbing – not when it's mature, not when every brand is doing it, not when it's starting to plateau. By the time a format feels proven to most people, you've already missed the arbitrage window. You're catching the end of the curve.

S – Scalable. If you can't replicate this format across accounts without manual effort, it's not useful for a distribution system. Great content that requires a genius to produce isn't a system – it's a dependency.

C – Convertible. This is the filter most people skip entirely. We run semantic analysis on the comments of any format we're considering doubling down on. If people aren't asking "what is this app," "where do I get this," "I need this" – it's not converting. Views and intent are not the same signal. If the comments are just reacting to the content, you have entertainment. You don't have distribution.

The VSC filter isn't about finding the perfect video. It's about not putting your system behind the wrong format at scale.

IV – the loop problem is an operations problem, not a creativity problem.

When we went beyond 10 accounts, I had a choice: hire more reviewers to stay in the loop, or build a system that didn't need me in it at all.

Hiring more reviewers is the wrong answer. It doesn't scale, and it means your content quality is now dependent on whoever you hired understanding your taste the way you do.

What I built instead was an AI playbook system with three components:

A – Brand context. What Flamme is, who it's for, what it sounds like, what it never does. Encoded once, referenced every time.

B – Examples. The actual pieces of content that hit the VSC criteria – the real proof of what working looks like for this brand and audience.

C – The prompt. The generation instruction that pulls A and B together and outputs at volume.

With this system, I'm not reviewing 30 pieces of content a day. I'm reviewing the system – checking whether the parameters are still accurate, whether the examples are still current, whether the conversion signal is still holding. I'm on the loop, not in it.

The output volume can scale without my time scaling with it. That's the only version of this that works.

V – the moat isn't the tool. it's the accumulated signal.

Every week, I see brands copying content we put out on Flamme. The format, the structure, sometimes the hook almost verbatim. It doesn't matter.

What they're copying is the output of the curve, not the insight that produced it. By the time content is visible enough to copy, it's already heading toward plateau. They're catching the tail.

The real compounding isn't in the content – it's in the data the system accumulates about what works. Which hooks produced comment-intent. Which formats drove Day 1 installs that stayed through Day 30. What the retention curve looks like for users acquired through product demos versus memes versus UGC.

Here's something I didn't expect: the hooks that work in content also tell you what works inside the product. The aha moment that makes someone stop scrolling is the same aha moment that makes someone stay in the app. We figured out parts of our onboarding because we could see what content was converting – and we reverse-engineered the emotional trigger back into the product. That cross-loop is only possible if your content system is generating signal, not just volume.

Nobody can buy that. They can copy the video. They can't copy the 18 months of iteration that produced the understanding behind it.

The mistake most brands are about to make is treating 550 videos a day as the goal. They'll crank the handle and watch the metrics go up – views, impressions, CPM trending down – and feel like they're winning.

The engineering teams who went to 4x code output and got 10% more shipped value made the same bet. Volume as progress. It looked like growth right until the review system collapsed under the weight of what it couldn't keep up with.

Distribution has always been expensive. AI just moved the expensive part to a place that's easier to ignore.

Build the system that compounds. Not the pile of content that churns.

- An