Image models can't write — so I write the text before they see it

The first time I went to generate App Store screenshots for one of my apps, I did what everyone does: describe the image I wanted to a model, headline included.

The render was gorgeous. The headline read “SHARF THE MENTAI LOAD”.

I rephrased, insisted, put the text in caps inside the prompt, tried another model. Always a letter dropped, an accent swallowed, a word invented. On a store listing the headline is the pitch — that’s not a cosmetic defect, it’s the product becoming unreadable.

Then I stopped asking the model to do something it cannot do.

The inversion

An image model doesn’t set type, it draws something shaped like type. That’s a difference in kind, not in quality: no future version fixes it reliably.

So the text stops going through it. A Python script draws it — exact font, computed size, pixel-placed — along with the device frame and the real capture inside. The result is a scaffold. The model then receives it with a two-block instruction:

• KEEP EXACTLY AS-IS: the text, word for word. The character. The colours.
• ENHANCE: blending, ornament, light, print-quality finish.

It no longer creates the image, it finishes it. And at that job it’s excellent.

KEEP EXACTLY AS-IS:
- The headline text: "SHARE" in terracotta, then "THE MENTAL LOAD,
  FINALLY SPLIT" in charcoal — same wording, same position, same
  approximate size, same colors

ENHANCE AND POLISH:
- Blend the mascot seamlessly into the background (remove the visible
  rectangular box behind him), keep his soft drop shadow
- Typography crisp and bold, professional print quality

The arithmetic that cost me a full set

Here’s the part I’d seen written nowhere, and which made me redo six screenshots.

Image models emit 9:16. Apple wants 1290 × 2796. Those are not the same proportions:

9:16                 = 0.5625
1290 / 2796          = 0.4614

The model’s image is therefore wider than the target. The crop takes the difference off the sides. Roughly 82% of the width survives — the rest is gone, silently.

My first set looked perfect on screen. After the crop, every headline had lost both its ends.

Hence a safe area set at 72%, not 82. The ten-point gap absorbs the model’s drift: however precisely you describe the frame, it adds an ornament that spills, nudges a badge, widens a composition. The script controls what it draws; the prompt has to repeat the constraint for what the model adds.

The bug I only found while packaging it

Turning these scripts into something shareable, I reread the function that fits the headline. It shrinks the font while the text exceeds the safe width, down to a 60px floor.

And if it still doesn’t fit at 60px? It returned the font anyway. The text overflowed in silence.

Which is to say: exactly the defect the safe area exists to prevent, reproduced by the guardrail itself. On a slightly long headline, the script would calmly produce a screenshot whose text was going to be cut off — without a word.

    font = ImageFont.truetype(path, 60)
    if draw.textlength(text, font=font) > max_width:
        print(f'WARNING: "{text}" overflows the safe width even at 60px — '
              f'it will be cropped. Shorten it or split it over two lines.',
              file=sys.stderr)
    return font

I only found it by running the script on a case I’d never had in my own app. That’s the strongest argument for packaging your tooling: you only really reread your own code when you hand it to someone else.

What it costs

You have to write the composer. Around a hundred lines of Pillow: headline fitting, device frame with bezel and rounded corners, placement. Not hard, not free, and not delegable to the model — that’s precisely the point.

Layout lives in code. Changing the composition means editing a script, not dragging a layer. For a set of six screenshots you redo every major release, that’s a win. For a one-off image, Figma is faster.

The model stays essential. The scaffold alone is correct and a little cold. Blending the character, the ornament, the finish: that’s the model. The idea isn’t to drop it, it’s to take away the one task it reliably fails.

What I take from it

The instinct, facing a model that fails, is to phrase the request better. Sometimes that’s right: the prompt was vague. But when the failure is always about the same thing — here, typography — it isn’t a phrasing problem, it’s a limit in kind.

The right answer isn’t a better prompt. It’s removing that task from the model’s scope, handing it to deterministic code, and leaving the model only what it does better than you.

The three scripts are public, in my skills repo: screenshot-scaffold. Composer, model call, crop — plus Apple’s exact sizes and the traps, in a separate reference.