AI Creative Pipeline for Growth

At Bestie, growth ads stopped being drawn one by one. I designed how the AI draws them.

Think of it as a design system whose only reader is a model.

A shipped ad, annotated with the rules the model could not break

I designed the image-generation framework, what each agent is for, what it must hand back and what it can never change, and wrote the prompts the image agents run on. I prototyped it with AI coding; engineering built it into production.

Every ad concept had to ship as a full set for the US, Japan, Europe and the Middle East, drawn by hand in Figma. One concept took about three days, and I was also designing the product itself. More hands wouldn't fix that; a different way of working would. I proposed to my director that the company go AI-native, starting with ad production, and the work was handed to me.

Asking a model to “make an ad” gave us exactly what you'd expect: images that looked like ads and nothing like Bestie. The worst failures repeated: a persona avatar stamped where the logo should be, and product screens the model simply invented. Because the pipeline runs on a daily schedule, a bad rule repeats its mistake every morning until someone fixes it.

That changed the brief. A model needs what a new junior designer needs: a clear job, a fixed structure to fill in, and rules it cannot break. So I split the work into separate jobs, gave each one a structure to hand back, and turned the brand guidelines into hard constraints.

These rules live in code behind PR review, not in the editable prompt, so no later prompt tweak can quietly drop them. The ad at the top of this page carries them.

The core decision was not to ask one prompt to do everything. A prompt told to play it safe and to explore at the same time gives you neither. So the image work splits into three jobs: iterate on what already wins, explore something new, and translate what works elsewhere in the category. A diagnose agent reads performance and feeds them.

Every agent reads the same thing first: my rules and its prompt. That is what decides what it makes, what it hands back, and what it is never allowed to touch.

After review, each generation is scored in market; winners seed the next round. Figures redacted.

Operators file requests on a board, by job and by market, and follow each one from “to design” to “done”. Every asset in the library carries a use_casetag, one of four, so the pipeline knows what kind of ad it belongs in. Outputs land in review folders by date, market and agent, where a person approves or rejects each image; approved ads upload to Meta paused. A prompt change can't be saved until it passes a dry run on a real winning ad.

Across the growth team over the same stretch, March to May 2026, monthly ad spend scaled about 3.9x while Meta CPI fell 41% and blended CPA fell 38%.