Anatomy of a Design Prompt

Most people write a prompt the way they'd describe a photo to a friend - one loose sentence. The model can parse that, but it fills every gap you left with its statistical average, and the average is exactly the bland, house-of-nobody image you didn't want. The problem isn't that the model is dumb; it's that you under-specified, so it defaulted. An architect would never hand a builder 'make it nice' and expect the right building - yet 'a beautiful modern house' is exactly that instruction to the model.

The fix is to think in slots. A strong design prompt names, in order: the subject (what the thing is), the style (the movement, period or named look), the medium (photograph, axonometric, gouache, V-Ray render), the composition (framing and viewpoint), the lighting (time of day, quality, direction), the camera (lens, angle, depth of field), and the materials (the specific finishes). You will rarely fill all seven, but you should decide each one consciously - because a slot you leave blank is not neutral. It is a variable the model sets for you, toward its mean.

So the discipline is simple and it changes everything: before you type, decide which slots matter for this image, and name them. The prompt becomes a brief you control instead of a wish you send off.

If you only get three slots right, make them these. The subject anchors the whole image: 'courtyard house' and 'apartment tower' send the denoiser to completely different regions before another word lands. Be concrete - 'a small single-storey courtyard house' beats 'a residence'.

Style is where architectural literacy pays off. 'Modern' is nearly meaningless - it was pinned to too many different pictures in training to steer anything. 'Kerala vernacular', 'Brazilian brutalism', 'Japandi', 'mid-century Palm Springs' each name a bounded look the model learned as a cluster. The more precisely you name the tradition, the more coherent the result. (Module 1.2 is entirely about which style words carry real weight.)

Medium decides whether you get a photograph, a hand drawing, a diagram or a 3D render - and it is the slot people forget most. 'architectural photograph', 'watercolour elevation', 'axonometric diagram', 'V-Ray interior render' each pull the image into a whole different production world. Naming the medium is often the single biggest jump in quality, because it tells the model what kind of picture this even is.

These three interact, and the interaction is where craft lives. A subject rendered as a 'pencil sketch' communicates a loose, early concept; the same subject as a 'photoreal dusk photograph' says 'this is resolved, this is real'. Choose the medium to match the stage you're at, not just the look you want - handing a client a photoreal hero of a scheme you haven't resolved sets an expectation you can't yet keep. The medium slot is quietly a communication decision as much as an aesthetic one.

These three slots borrow the language of photography and cinematography, and they work because the model learned from captioned photographs that used exactly those words.

Composition places the viewer: 'wide establishing shot', 'eye-level street view', 'aerial three-quarter view', 'tight interior corner'. Without it you get the model's default - usually a slightly dull frontal hero shot.

Lighting does more emotional work than any other slot. 'warm golden-hour light', 'flat overcast light', 'dramatic raking side light', 'soft north light through a window' - each sets mood, reads material differently, and casts the shadows that give a building depth. Interiors especially live or die on this slot.

Camera is the finish: '24mm wide lens', 'tilt-shift', 'shallow depth of field, f1.8', '35mm architectural, f8'. Lens terms change proportion and drama; a wide lens exaggerates space, a longer lens flattens and dignifies. You are not taking a real photo, but the model behaves as if you were - so speak to it in a photographer's terms and it obliges.

The reason this trio works so reliably is worth holding onto: the model learned overwhelmingly from captioned photographs, and photographers caption with exactly these words. When you write 'raking side light, tilt-shift, eye-level', you're not inventing a language the model has to interpret - you're speaking its native tongue, the vocabulary its training data already used for millions of images. Borrow the discipline of a photographer briefing a shoot and you'll out-prompt someone twice as fluent who never thought to name the shot.

The last slot, materials, is what separates a generic surface from a specific building: 'laterite and exposed concrete', 'travertine, oak and brushed brass', 'white render with teak louvres'. Interiors and facades reward this slot heavily - it is where a room stops looking like a stock render and starts looking like your scheme. Name two or three finishes, not ten; a short, decisive material list reads as a considered palette, while a long one muddies into visual noise the model can't resolve.

Now the part most people miss: word order is itself a control. These models weight terms roughly by position - what comes first pulls harder, and the tail of a long prompt is close to a whisper. So front-load the one thing that must be true. If the building's brutalism is non-negotiable, 'brutalist chapel...' leads; if the material is the whole point, lead with it. Bury the nice-to-haves at the end where a little drift is fine. This is why two designers can type the same seven words and get different images: the one who put the load-bearing term first got the building they meant.

Group by slot with commas, too - the model reads comma-separated clauses as distinct instructions, so 'brutalist chapel, board-formed concrete, raking evening light, 24mm, backlit' is clearer than the same words run together in a sentence. Same words, reordered, produce a different image. Structure is not tidiness; it is steering. That is the whole reason this lesson exists before the vocabulary lesson - once you can see the slots, every new word you learn has a place to go.

Watch the difference on one example. A beginner writes: 'a beautiful modern Indian house, very nice, high quality, amazing design'. Every word there is either a taste-adjective the model can't act on or a term ('modern', 'Indian house') so broad it collapses to a stereotype. The result is predictable - a glossy, generic, nobody's-house render.

Now the same intent as a slotted brief: subject 'a single-storey courtyard house'; style 'Kerala vernacular'; medium 'architectural photograph'; composition 'wide eye-level view from the entry court'; lighting 'soft monsoon-overcast light'; camera '24mm, f8'; materials 'laterite walls, Mangalore-tile roof, teak columns'. Strung together and comma-grouped, front-loading the non-negotiable: Kerala vernacular courtyard house, laterite walls, Mangalore-tile roof, teak columns, wide eye-level view from the entry court, soft monsoon-overcast light, architectural photograph, 24mm f8.

Nothing in that prompt is flowery, and that is the point - it reads like a project brief because it is one. Every clause is a decision you can defend and, later, a variable you can change one at a time. Notice too that the weak version was longer in spirit (all those empty superlatives) yet said less; the strong version is dense with named things and says far more. This is the habit to build: think in slots, name the ones that matter concretely, order them by importance, and stop. Do that and you have already left the crowd of people typing wishes into a box.