Getting the words right is no longer about spelling

You generate a design and three of the four words are perfect. The fourth has a letter that's collapsed into the one beside it. So you ask again — and this time the fourth word is clean and the second one has drifted somewhere you never asked it to go.

Anyone who has put type on a shirt with these tools has had that afternoon. It feels like the model is being contrary, and it isn't: nothing carries over between two images except your prompt. Every attempt is a fresh roll, not an edit of the previous one.

Which is the thing actually worth knowing, because the problem everybody still braces for — will it even spell it — stopped being real a while ago.

Can image models spell now?

Yes, and not only short punchy words. Long words, full lines of copy, distressed and chrome-bevelled and non-Latin lettering all come back correct.

What hasn't gotten easy is control. You can't nudge kerning by asking, you can't move one word two centimetres, and the model has no idea which of your words is the important one. So the difficulty moved: from "will it spell it" to "will it put the right word in the right place at the right weight".

That's a much better problem to have, and it's mostly not yours.

Do I need to learn any of this?

No — and that answer goes before the craft below rather than after it.

Say what you want it to say and how it should feel. What comes back is expanded into full directions: every word quoted so its spelling locks, every word given its own position so nothing repeats itself, the lettering described as a material rather than a mood. Those are the three things that go wrong, and they're handled.

The rest of this is what that expansion is doing on your behalf. It's worth reading if you want to take the wheel and push a design somewhere specific — not because the shirt requires it of you.

What locks the spelling and the layout

Two habits, and they fix two different failures.

Each word gets quoted on its own — the word 'NIGHTMARE', rather than a description of the phrase it lives in. Quoting the individual string is what makes the characters the content instead of a vibe to be interpreted.

Each word gets its own position — "'COLD' arched across the top, 'BLOOD' set beneath the figure." Without a position each, the classic failure is a word turning up twice: once where you wanted it, once as decoration on the other side of the image.

What about long words, or a lot of them?

There used to be a third rule here — keep words short, no more than two per design. That was a spelling limit and it's gone. What's left is a layout judgement, which is a different thing entirely: the reason to use two big words is that someone glancing at your chest from across a room won't read three, not that three won't render. A design that genuinely wants a long word or a full line of small type can have one.

Why doesn't the font ever look how I asked?

Because "bold, aggressive, powerful lettering" is four adjectives, and nothing renders an adjective.

The single biggest upgrade available is naming a material and a process instead of a mood. Compare that phrase with:

• molten liquid chrome dripping blood-red
• jagged blade-like letterforms fused together
• polished silver chrome bevel with a blood-red inner glow
• heavy condensed uppercase stencil, ink-stamped and partly faded

Every one of those names a substance and something that happened to it. That's a thing to draw. "Aggressive" isn't.

The same principle runs through everything else: expressiveness comes from concrete nouns, not adjectives. When a design isn't hitting hard enough, what fixes it is another object — a snapped chain, wisps of smoke, letters cracking apart — never another adjective.

The hex code trap

Colour codes don't belong in a prompt, which surprises people. Models don't render by hex value, so #B00020 is false precision — you don't get that red, and every so often the code leaks into the artwork as stray characters, which is an annoying way to lose an image. Perceptual language does the job: flat near-white, zero texture, deep blood red, cool bone. Hex stays where it belongs, as a value in a spec for whatever produces the physical piece.

Can the letters be deliberately mangled?

Yes — and this is the part that's quietly gotten very good.

The bootleg tee tradition has never used clean type. Letterforms get fused into each other, bowed into arcs, spiked, melted, cracked apart, warped until they're barely words. Models render that deformation well now — arguably more reliably than they render clean type, because in clean type a slightly wrong curve is a defect, and here it's the point.

How it gets asked for

Two moves, and the first one matters more:

• The whole image gets anchored to the tradition, not just the font — 90s bootleg rap tee graphic, 1980s heavy metal album cover art. The anchor sets the ceiling on how wild the lettering is allowed to get, and nothing else moves that ceiling as far.
• Then the deformation is described as a process — jagged blade-like band-logo letterforms fused together, bowed downward like a bridge, razor spiked serifs.

You're trading legibility for ferocity on purpose here, and that trade is the genre. The only thing worth being deliberate about is that it was a trade — a word nobody can read should be a choice you made, not one that happened to you.

Should the text be part of the artwork at all?

Not always, and knowing which is which is most of the craft.

Generated type is right when you're exploring, when the words are few and bold, and when the letters should belong to the same world as the image — chrome that matches the chrome, distress that matches the distress. It looks made rather than assembled.

It's the wrong tool when you need a precise, repeatable layout: an exact font, exact kerning, the same lockup across twenty designs, lowercase script with punctuation. You can't converge on that by re-prompting, for the reason this article opened with — every generation is a fresh roll. That's when type gets set properly and laid over the art instead.

The tell is simple: are you exploring, or executing a spec?

What about phrases that are already someone's brand?

Worth checking, because gym-and-motto language collides with trademarks constantly.

The wording that sounds most natural in a category is usually the wording somebody already built a company on. We've had to swap words at the last step for exactly this: a design reading IRON WILL became STILL STANDING, another NO LEASH became UNLEASHED — not because the phrases were unusable in speech, but because they were somebody's name in apparel.

Short, generic-sounding, category-defining phrases are the risky ones. Something specific to you — your own words, your own joke, your own number — is both safer and better. (Our content policy covers where the line is.)

How do I actually make one?

Describe it on JustOG in a sentence, or write the prompt yourself if you'd rather steer. Either way you see the result on the real garment before anything is made.

Read the words one last time before you order. That part never gets automated, and it shouldn't.