AI’s biggest illusion is making you believe it actually understood you.

Today I got wrecked again by this whole “prompt engineering” thing.

I spent a day and a half trying to lock down the visual details of an AI-generated image —

trying to get a consistent style that actually feels like my own.

I wrote everything out:

font size, typeface, color codes, layout, composition…

super detailed, super professional.

And the AI kept responding enthusiastically:

“I got it.”

“I understand.”

“I’ll follow your settings.”

Even throwing out suggestions like it was fully on board.

But after round after round of adjustments, you eventually realize:

It never understood any of it — and it never followed a single detail you gave.

What makes it worse is the hallucination part.

It keeps feeding you this fake hope, this feeling of

“Okay, almost there, one more tweak…”

So you keep trying, keep refining, keep adjusting.

And then it hits you:

every minute you spent was a complete waste.

While you thought it was cooperating,

it was actually just flattering you while ignoring every single detail you described.

What completely broke me was discovering this later:

the model that generates the image and the model that handles layout/typography aren’t even the same model.

So expecting it to understand both “visual content + font layout”

with one set of prompts

is literally asking it to do something impossible.

That’s why trying to control all these details with prompts just doesn’t work.

But the moment you give the AI a reference image you actually like

and tell it, “Match this style,”

the results are instantly ten, twenty, fifty times better.

At this point, my only conclusion is:

stop overthinking prompts — just use an image reference.