AI Image Prompts

Same brief, same model, same afternoon. The only difference is where the prompt came from.

The left-hand image gets three things wrong that nobody asked it to do. It came back square when the brief said vertical. It wrote its own headline — “Script to a creative odyssey” — which was never requested, and set it across the colour-grading panel. And every image inside it is a stock cliché of filmmaking: typewriter, film set, grading desk, cinema screen. None of it says anything about the product.

The right-hand image is the same brief run through a structure from the library: four stacked panels, a shot size written for each panel, and a technical tail pinning down lens feel, lighting, depth of field and grade. Correct ratio, no invented text, and the sketch in panel three reappears as the finished shot on the monitor in panel four — a continuity we asked for and the model held.

What it is

A cookbook, not a generator. Roughly 14,500 image prompts that have already been used in public, sorted into eleven categories, each entry carrying its full prompt text, a sample image, and whether it needs a reference image. You go and find one that already has proof it works, and take it.

What we found running it

Search by layout, not by subject. This is the finding that matters. Searching our actual subject returned portrait and fashion editorial almost exclusively — Paris street style, red dresses, actresses — and nothing resembling B2B product marketing. By topic, the library was useless to us. What was worth having was one entry’s skeleton, which came from a prompt about a woman on a Nordic coastline and transferred to our subject without complaint.

The sample image is the best case, not your case. We also ran one library prompt completely unchanged and compared it to the sample published alongside it. The content reproduced faithfully — same four shot sizes, same wardrobe, same composition. The layout did not: our run drew literal film sprocket holes and stamped FRAME 1 / 2 / 3 labels across the image, printed FRAME 3 twice, and pushed FRAME 4 off the canvas. The sample has none of that. Those prompts were written against a different image model than the one we ran them on; the descriptive layer crossed over, the structural instructions did not.

Its own search command will blow up your context. The skill says, in capitals, never to load a category file — and then hands you a grep with no result limit. We measured it: searching one common word returned 4,645 lines and 4.85MB, on the order of 1.2 million tokens. That is a saving of less than 3× over the whole-file load it tells you to avoid. A small search script that returns ids and titles answers the same query in about 370 tokens.

• Counts disagree with themselves — the real number of unique entries is 14,546; the manifest claims 15,089 and the documentation says 14,398. The per-category numbers are correct; only the two totals are wrong.
• It phones home on every invocation — a setup check that refreshes from GitHub if the local copy is over a day old. We read those 117 lines: it fetches and writes files, and does not execute anything it downloads.
• There are real people in it — some entries are built around named public figures. If you are piping this into a video model that refuses photorealistic real faces, you need a filter in between.

Four tasks, each run with the skill and again with skills disabled, identical wording both times. Both arms always returned usable prompts. What separates them is where the prompts come from.

The skill arm treats this as retrieval. Task 2 asked it to find a few style templates from the library, then adapt the one I pick; it went to the prompt library, named the template it selected (a business-style flow diagram), and saved the reference it worked from as reference-style-template.jpg (64,329 B) alongside the prompt document. The control arm reported the directory was empty, invented three SVG templates of its own, and adapted one of those. Both returned a cover design — only one answered the question that was asked. Task 4 has the same shape: the skill arm shipped seven files, six of them real 16-bit reference images from the library (97 KB–227 KB each) next to pixel-rpg-character-portrait-prompts.md; the control arm shipped the markdown alone.

The retrieval also fails honestly. In task 3 the skill arm reported that the library held no template matching a three-laptop comparison thumbnail, wrote the six prompts from scratch, and said so in its answer rather than implying a match it did not have.

That retrieval is what you pay for. On the three straightforward briefs (tasks 1, 3, 4) the skill arm cost $0.36 / $0.43 / $0.49 and took 15 / 16 / 20 turns, against $0.13 / $0.14 / $0.15 and 5 / 5 / 6 turns without it. On task 2, where both arms had to work, the costs were comparable ($0.48 with, $0.58 without).

What we did not test: image quality. No images were generated in either arm, so we have no evidence that the retrieved prompts render better than the invented ones — only that they are traceable to a source. The library is external; if it is unreachable, the skill degrades to roughly what the control arm did.

Related

For camera movement rather than image composition, see Seedance 2.0 Prompting. For deciding what the shots should be at all, see Video Spec Builder.

Installing

Not ours — install from the source repository. MIT, and free to run: it holds prompts, it does not call a model. Generating anything from them costs whatever your image provider charges.