A diagram, in this house, is a combination of attractive 2D or 3D forms that illustrates a concept before any word is read. It is not the engineer's box-and-arrow figure: it is composition, material and light saying the idea.
Method: skill concept-diagrams. Engine research: image models for skills.
Every diagram is one pick from each bank: a structure that says the concept, dressed in a material that gives it beauty. The banks are open: every approved diagram adds its pair. There may be many other subjects; this is the start.
A material can be photoreal (the Earth, a delta) or crafted (clay, paper). Both are welcome; childish is not.
The structure is the argument: pick the one whose geometry IS the concept.
Eight studies made on day one, all with the cheap default engine, centimes in total. The first three play the same concept (an institutional world above, a living territory below, one new bridge) in three materials; the other five each play a different concept. No image carries a single word: real text always sits on the page, above.








One family, four recipes. The picture under each is a real thing it made.
The house kind, above: forms that carry an idea before a word is read. Made with the cheap image engines, the text laid over the picture afterwards, never inside it.
The engineer kind, twenty-seven sorts: architecture, flow, sequence, timeline, quadrant. One self-contained page, one ink, one accent.
The same subjects, but the reader hovers to trace a route and the map can be checked against a real repository. It can move, and it exports a picture or a film.
Machines, stores and documents as blocks seen from the corner, with a built-in icon set. Driven headlessly to a picture, so it lands in a page or a slide.
Looked at and refused, so nobody proposes them again. The whiteboard helpers (Miro, FigJam, Lucid) and the slide-visual makers need a human at the mouse, which fails the one test the studio applies: we must drive it end to end. arch-flows-visualizer makes the same self-contained page as the two above from five commits and no real licence. Datadef is the only one worth remembering: it exposes a machine interface and puts real product icons on an architecture, so it is a candidate the day a page needs vendor icons that FossFLOW does not carry.
| Engine | Use it for | Cost | State |
|---|---|---|---|
| Gemini image API (gemini-3-pro-image) | The default: every study on this page. Photoreal and crafted materials both. | Fractions of a centime per image | Live, key in the keys file, proven pattern in the skill |
| GPT Image 2 (OpenAI API) | Complex compositions; keeping one style across a family of pages (16 reference images + mask editing) | ~$0.05 per image, pay as you go | Waiting for top-up |
| Recraft V4.1 | Editable vector SVG shapes; a trainable house style | $0.08–0.30 per vector | Waiting for top-up |
| FLUX.2 on RunPod | Batches and upscales at open-model prices | ~1.3 centimes per job | Live in the video toolkit |