I used to think the hard part of making good AI graphics was the prompt.

You know the ritual.

You open ChatGPT, describe a visual in increasingly desperate paragraphs, add “make it more premium”, add “less corporate”, add “more editorial”, and suddenly you are negotiating with a machine like it is a very talented intern who also had three espressos and no briefing document whatsoever.

And sometimes it works.

But most of the time you get something that looks almost good.

Which is the worst category, because “almost good” means you will now spend 37 minutes pretending the weird spacing is a style choice.

So I changed the workflow.

The image model is not the designer.

It is the asset generator.

The layout system is the designer.

My current process looks roughly like this:

  1. Write the actual argument first.
  2. Convert it into a visual system.
  3. Generate weird, specific image assets with AI.
  4. Place them into HTML and CSS.
  5. Render the final graphic deterministically.
  6. Check if it survives the phone feed.

That last step is where many AI graphics go to die, by the way.

They look wonderful at full screen.

Then LinkedIn turns them into a small rectangle between someone’s promotion announcement and a post about “strategic alignment”, and suddenly your beautiful micro-labels are decorative dust.

Give the model the chaos job

The important distinction is this:

AI is very good at creative variation.

HTML is very good at structure.

So I do not ask AI to be structure. I ask it to create the strange raw material: a paper factory of prompt fragments, a glowing terminal fossil, an impossible editorial still life of browser windows and copper wires.

Then HTML does what HTML is boringly excellent at:

Alignment. Hierarchy. Spacing. Repeatability.

The boring part is the leverage.

Once the system exists, I can generate ten different graphics without rebuilding the whole thing from scratch. I can swap the image, adjust the copy, rerender, compare, and keep the version that actually works.

This is the bit people miss with AI.

The magic is not “one prompt creates one perfect thing”.

That is the just pay shipping version of AI creativity.

The useful version is building a small machine where AI creates options, code enforces taste, and the human decides what deserves to survive.

Which is, annoyingly, still work.

But at least now the work compounds.