Your post-purchase email is hurting retention

You have a raw product image (hint water, in this example):

You generate an AI image and it’s… usable.

The context makes sense.

The product looks real.

The background isn’t doing anything stupid.

But one thing stands out. The lighting.

Not awful. Just not right either.

So you do what any normal founder or designer would do.

You ask for a small tweak, such as, “Can we fix the lighting a bit?”

That’s usually when everything starts drifting.

At first glance, this feels like a lighting problem.

It isn’t, but it’s a control problem.

AI has no idea which parts of an image you’ve mentally approved and which parts you’re still flexible on. When you prompt in normal, conversational language, the model treats the entire request as open for reinterpretation.

So even when you intend to fix just one thing, the model rethinks:

That’s why “small tweaks” often cause surprisingly big changes.

This wasn’t a “JSON prompting is cool” experiment.

It was a simpler, more practical question:

Does the way you structure a prompt change how much freedom the model gives itself before iteration even begins?

To keep things fair, we held everything else constant:

The only variable was how the prompt was written.

One version used normal, natural language.

One version used structured JSON.

Before looking at the outputs, one thing needs to be clear.

These were two clean generations, run in separate tabs.

They were not follow-up edits on the same image.

That’s why you’ll notice differences beyond lighting:

And that’s intentional.

This test is not about pixel-matching.

It’s about how much interpretive freedom the model takes upfront.

Natural Language Prompting

This output gets most of the way there. The product feels real, the setting makes sense, and it’s something you could plausibly refine further.

The issue is the lighting. It’s slightly off, and fixing that would naturally be your next prompt.

But nothing else is protected. So when you try to fix lighting, you’re also risking changes to framing, background, and scale. This is where iteration usually starts to feel messy.

This output isn’t “perfect lighting” either, but much better than the previous one.

What’s different is behavior.

The scene feels more contained. The model makes fewer surprising choices. Even though elements differ from the other image, the overall result stays closer to the original intent.

This image isn’t more creative, but is more predictable.

This isn’t about quality, but most importantly, the interpretive spread.

Natural language invites the model to explore.

Structured prompts narrow the space it explores within.

JSON doesn’t freeze layouts or magically lock scenes. What it does is make your decisions explicit, so the model spends less energy reinventing them.

That difference shows up before you even start iterating.

The paid section below breaks down:

To Continue Reading..

$17/Month. That’s ~$2/issue.

7 days free trial. Cancel anytime.