I ran 461 prompts from the popular collections. This is what came out.

There is a genre of repository that does very well on GitHub: the prompt collection. A few hundred prompts, a screenshot next to each one, tens of thousands of stars. I started from two of them — both properly licensed, both excellent — intending to translate, tidy and republish.

Before publishing anything I ran all of them. That is the whole of the method, and it is not clever: fire each prompt at gpt-image-2, keep the image that comes back, and look at it. It cost about eighty dollars.

What follows is what that turned up. None of it is a criticism of the source collections, which were written for a different audience and are honest about it. It is what nobody sees because nobody runs them.

1. Twenty-five images came out in a language nobody asked for

The catalogue is meant for English and Spanish speakers. Twenty-five of the generated images had Chinese, Japanese or Korean lettering in them — headlines, buttons, price tags, shop signs.

The obvious explanation is that the prompts were Chinese. That was true of some, and translating those was easy. It did not fix the problem, and chasing it down took four separate attempts, each of which felt like the last one.

1. The prompt body had CJK characters in it. Caught by looking for the characters. Ten prompts.
2. The prompt said, in English, to render the text in Chinese — "written in Simplified Chinese". Not one CJK character in the file, so the first pass sailed past it. Twenty-five more.
3. The variable examples. This is the one that mattered: forty-five entries had example values in Chinese or Japanese, and the pipeline was picking the first example when it generated the preview. The prompt was clean English and the image still came out in Japanese, because we had told it to.
4. And the model's own bias. With an English prompt and no CJK anywhere, a green tea bottle still gets a Japanese label.

2. Instructing the model is not a guarantee

After translating everything I added an explicit instruction: all lettering in English, no CJK anywhere in the frame, not even as background detail, not even when the scene is set in Asia. Then I regenerated the twenty-five.

Twenty-five images came back. Several still had Japanese text in them.

A second pass recovered twenty-two of them. But that is luck, not a method. When the scene is a green tea bottle or an anime poster, the model pulls toward the script that belongs to it, and a sentence in the prompt does not reliably stop it.

So the guarantee had to move. Every image now goes through a second model that reads the text inside it and reports what it sees. If it finds lettering in a script the prompt did not ask for, or a real company logo nobody declared, the entry does not publish. Three are held back right now — they exist, they are paid for, and they are not on the site.

That check had been running for days before I wired it to anything. It was detecting the Chinese correctly and writing the result to a file nobody read. The detector was not the hard part; connecting it to the publish step was.

3. A brand detector does not see a franchise

Two entries turned out to reproduce the twelve Gold Saints from Saint Seiya, with their names printed under each portrait. One of them was on the front page.

The check that should have caught it looks for company trademarks — Nike, Apple, Coca-Cola — and a franchise is not a company. The prompt said "the 12 Golden Saints from [FRANCHISE]", which reads like a neat parameterisation until you see what comes out of it. Both are out of the catalogue.

A sweep found fifteen more entries that name someone else's intellectual property. Most are style references — "in the style of a 3D animated film" — which is common practice and a different thing. The line I settled on: naming a look is fine, naming the characters and reproducing them is not.

4. The people in the images were not who I expected

When a prompt describes someone without saying what they look like, the model decides. I measured every image in the catalogue — how many people, how prominent, apparent origin — rather than guessing.

Unclear or not visible

28%

White / European

25%

East Asian

21%

Latino / Hispanic

20%

Black / African

3%

Middle Eastern

2%

That is after correcting it. Before, one in every four or five faces in a catalogue aimed at Europe and Latin America was East Asian, and the cause was not the source of the prompts. It was our own variable examples again: when a prompt said [MODEL_DESCRIPTION] and the first example was an East Asian woman, that is what the catalogue filled with.

The correction is not swapping one ethnicity for another, which just moves the bias. An appearance profile rotates across cases so the catalogue as a whole varies. Black and African representation is still at three per cent, which an equal rotation does not fix and I have not solved.

5. Half the before/after comparisons compared nothing

Thirty-nine prompts said they start from a photo you supply. I generated a before/after for each, fed with an image from our own catalogue, and put them behind a slider.

Then I looked at the pairs. In more than half, the two halves had nothing to do with each other: a cosmetics jar on the left, a penguin-shaped gift box on the right. Not a pairing bug — those prompts say "design a packaging sheet", not "transform this image", so the model ignores the input and makes something new.

An algorithm can check that an image has no foreign brand, no CJK and the right aspect ratio. It cannot check that the after has anything to do with the before. Twenty-one of the thirty-nine turned out to be real transformations. Those are the ones with a slider now.

What this is worth

Nothing here is a clever technique. It is what happens when you do the boring thing — run the prompt, look at the image — and it turned up problems that no amount of reading the prompt would have found. Four of the five could only be seen in the output.

The catalogue is at promptfeed.app, free, no account, and the prompts are on GitHub under MIT. If you find one that no longer works, that is the most useful bug report there is.