The "Compositing Engine" (The Winner)

Hey everyone! 👋

I wanted to share a massive "lesson learned" from a project I just pivoted. I was building an AI Thumbnail Generator for Inside Success TV, and halfway through, I scrapped the whole thing.

Not because it didn't work but because the design was wrong. If you’re building AI tools for your portfolios or clubs, this might save you a lot of time and money.

I stopped asking "Can AI do this?" and started asking "What part actually needs AI?"

Now, instead of one big AI generation, I use a modular pipeline:

• Real Assets: High-quality stock photography from Pexels.

• Engineering: Node.js and Sharp handle background removal and layering (the "sandwich method" where text sits behind the subject for 3D depth).

• The AI Role: Claude does one specific job translating a simple prompt like "Baywatch episode" into precise instructions for the machine.

• The Result: Cost dropped from $0.65 to $0.003 per session, and the quality matches exactly what a human designer would build.

💡 The Big Takeaway

The lesson isn't that AI is bad. It’s about leverage.

AI at 85% of your pipeline is a liability. At 10%, it’s leverage.

7 comments