The third prompt in the collection, and the first one whose output is not meant to be looked at.
The first and second were prompts for a photograph — you run them, you get a picture, the picture is the point. This one produces a character sheet: twenty-one panels of the same invented woman on a grey studio background, five body angles across the top, eight expressions through the middle, eight poses along the bottom. It is not a picture. It is a reference asset, built to be fed into a video generator so that whatever gets made next has one consistent person in it.
She has a name — Linh Anh — and that is the first clue about what this is actually for. Nobody names a photograph.
Look at where the words go and the subject is obvious: consistency. There is a section headed Core visual identity, a second headed Important consistency rules that mostly restates it, and then a negative prompt longer than either of the first two prompts in this collection were in full — no inconsistent face, no changing hairstyle, no changing body proportions, no duplicate faces. Roughly a third of this prompt is one instruction said four different ways: keep her the same.
Which tells you what the hard problem is. Nobody writes four hundred words defending against a failure that does not happen.
And it half works. The outfit holds across all twenty-one panels: the champagne satin top, the black high-waisted trousers, the small gold earrings, the thin bracelet, consistent down to the sheen on the fabric. The expression row delivers exactly the eight expressions asked for, in the order asked for. The pose row delivers exactly the eight poses, in order. For a single generation that is a lot of instruction followed correctly.
What drifts is her hair — which is precisely the thing the prompt forbids twice. It is specified as straight black shoulder-length, and in the close-up row it is shoulder-length. In the rear and three-quarter-back views it falls well past the shoulder blades. Naming a failure in the negative prompt does not prevent the failure. It only records that the author already thought it was likely.
The layout is the other soft spot. The top row asked for five full-body angles in a clean grid and came back ragged, with a blank panel sitting in the middle of the row for no reason at all. Clear readable layout and clean composition are both in there; the model read them as mood rather than as instruction. Composition is not something you get by asking politely.
One piece of housekeeping for this collection going forward. This one came out of Gemini — the filename says so, which is more than I can tell you about the first two. The prompt above is not portable, because half of what makes it work is how one specific model reads one specific phrase. A library of prompts with no record of what produced them is a library of guesses, so I will be noting the model from here on.
