Architecture Prompt Generator: Build Design Prompts That Work

TL;DR: You are probably here about buildings, not about AI job titles. An architecture prompt generator is a fill-in-the-blanks block with eight slots: project type, site, programme, massing intent, materials, mood, camera and output medium. Fill them, concatenate, paste. It produces renders and concepts, never construction drawings.

What is an architecture prompt generator?

An architecture prompt generator is a structured block of text you fill in and paste into an AI image model to get architectural visuals. It is a template with slots, not a piece of software, and nothing needs installing for it to work.

The name is doing two jobs on the internet right now, so let us clear that up before anything else. If you design buildings, you want prompts that produce massing studies, facade options and presentation imagery. If you work in AI, "prompt architecture" means the internal structure of an instruction you send to a language model. This page is for the first group. Our company being called Prompt Architects is a coincidence that sends a lot of practising architects here, and it seemed rude to keep serving them the wrong article.

The reason to build a block rather than collect prompts is that architectural briefs are all different and prompt lists are all the same. A list gives you somebody else's cultural centre in Lisbon. A block gives you your own mixed-use scheme on a sloping corner site in Manchester, with the north light you actually get. If you do want a curated set to raid, we keep one at AI prompts for architects. This page is the machine that produces them.

The architecture prompt generator (copy this block)

Fill in the eight slots. Delete the labels. Read the result back as one sentence. That is the prompt.

[OUTPUT MEDIUM] of a [PROJECT TYPE], [SCALE / STOREYS],
on [SITE: location, context, orientation, topography],
containing [PROGRAMME: the two or three spaces that define it],
[MASSING INTENT: the formal move in plain words],
[MATERIALS: envelope, then structure, then one detail],
[MOOD: time of day, weather, season, quality of light, people or none],
[CAMERA: view type, height, lens],
[MODEL TAIL: aspect ratio, style controls, exclusions]

Assembled, a filled block reads like this:

Photoreal architectural render of a four-storey community library,
on a tight urban corner site in Manchester, red brick terraces
opposite, north-facing entrance, ground falls 2m east to west,
containing a double-height reading room, a basement archive and a
street-level cafe, a simple brick bar with the reading room carved
out of the upper two floors as a recessed loggia, dark red brick
with precast concrete lintels and bronze-anodised window reveals,
overcast October afternoon, wet pavement, a few people entering,
eye-level street view from the opposite pavement, 35mm lens,
--ar 3:2 --s 250 --raw --no cars

That took under a minute to write, and every clause in it is a decision you would have had to make anyway. The value of the block is not speed. It is that it forces the eight decisions to the surface instead of letting the model make them for you, which it will, badly and identically, every time.

What goes in each slot?

Each slot has a small set of values that reliably change the output. Pick from the banks below, or write your own in the same grammar.

Project type and scale. Be specific about typology and storey count. "Building" produces a glass box. "Four-storey community library" produces a library. Add the structural system if you care about it, because "CLT-framed" reads very differently from "concrete-framed" in the output.

Site. This is the slot that separates a real study from a stock image. Give location, immediate context, orientation and topography. Models do not know your site, and they will invent a flat plot with generic neighbours unless you describe what is actually there. Orientation matters more than people expect, because it constrains where the light can come from and stops the model defaulting to a sunset behind the entrance.

Programme. Two or three defining spaces, not a schedule of accommodation. The model cannot lay out a building, but naming the double-height reading room tells it something about section that "library" alone does not.

Massing intent. The formal move, in plain words.

Materials. Name the envelope first, the structure second, and one detail third. Three material terms is the sweet spot. Five turns into mush. Reliable vocabulary: board-formed concrete, weathering steel, charred timber cladding, exposed CLT, rammed earth, glazed white brick, terracotta baguettes, anodised aluminium, fritted glass, perforated brass mesh, travertine, exposed blockwork, standing-seam zinc.

Mood. Time of day, weather, season, light quality, and whether there are people. This is where most prompts collapse into the same golden-hour render. Alternatives that read as architecture rather than advertising: flat overcast north light, low winter sun raking across a facade, blue hour with interiors lit, morning fog softening the background, wet pavement after rain, harsh midday summer with hard shadows.

Camera. Say the view type, the height and the lens.

Output medium. The slot people forget, and the one that changes the result most. "Photoreal architectural render" is one option out of many: white chipboard massing model, basswood study model on a site model, hand-drawn ink and watercolour, technical axonometric line drawing, section perspective with poché, competition board collage with cut-out figures, diagrammatic exploded axonometric.

Worked example: concept massing

Massing is where the generator pays for itself, because the output medium slot lets you ask for a study model instead of a finished building. A model photograph reads as a question. A photoreal render reads as an answer, and clients respond to it accordingly.

White chipboard massing model of a 12-unit residential scheme,
three storeys, on a narrow south-facing plot between two Victorian
terraces, ground falls gently to the rear, containing duplex units
over a shared entrance hall, three linked volumes stepping down
the slope with a glazed link between each, uniform white card with
laser-cut window openings, no material texture, photographed on a
white site model under diffuse studio light, three-quarter aerial
view from 30 degrees, 90mm tilt-shift, shallow depth of field,
--ar 4:3 --s 50 --raw

Note the low stylize value. Midjourney's own documentation describes stylize as a slider between literal prompt adherence and its house style, with a default of 100 on a 0 to 1000 range (checked at docs.midjourney.com on August 26, 2026). For massing you want literal. For a competition board you might not.

Generate four, print them, and put them on a wall next to the two options you drew yourself. The useful outcome is rarely that the AI option wins. It is that seeing eight bad variants clarifies which of your two is right, and why.

Worked example: materials and facade

Facade studies are the one place where holding everything else constant genuinely works, because you are testing a single variable against a fixed frame.

Photoreal close-up of the facade of a four-storey office building,
urban street frontage, north-facing, containing open-plan floors
above a double-height lobby, a regular structural bay with deep
reveals, [MATERIAL], flat overcast light, no people, straight-on
elevation view from across the street, 50mm lens,
--ar 16:9 --s 100 --raw --seed 1234

Run that block four times, substituting only the material slot: dark glazed brick with recessed mortar; weathering steel panels with exposed fixings; vertical terracotta baguettes over grey glazing; board-formed concrete with a shadow gap at each lift. Everything else is frozen, including the seed, so the differences you see are the material.

Two warnings. The model will invent a structural rhythm that does not correspond to any grid you have designed, and it will produce beautiful junctions that cannot be built. Treat the output as a mood reference for a materials conversation, not as a detail. If you find an image online that has the quality you want, reverse-engineering that image into a prompt is usually faster than describing it from scratch.

Worked example: interior and spatial atmosphere

Interiors need the camera and mood slots carrying most of the weight. Materials matter less than light does.

Interior architectural render of the main reading room in a
community library, double-height space, on the upper two floors
of a brick corner building, containing long shared tables and
perimeter stacks, a recessed loggia bringing daylight deep into
the plan from the north, exposed CLT ceiling with acoustic felt
baffles, pale terrazzo floor, blackened steel balustrades,
flat overcast winter daylight with warm task lighting on the
tables, a dozen people reading, one-point perspective from the
entrance at standing eye height, 24mm lens,
--ar 3:2 --s 200

The --no parameter is worth knowing here. Midjourney's documentation gives --no fruit, apple, pear as a valid comma-separated example, and warns that its moderation reads every word after --no independently, so --no modern clothing is parsed as "no modern" and "no clothing" (docs.midjourney.com, checked August 26, 2026). For interiors, --no plants and --no signage clear out a lot of default clutter.

Gemini handles this category differently because you can hand it a photograph of the existing space and prompt against what it sees. We covered that workflow separately in architecture prompts for Gemini.

Worked example: presentation and competition boards

Boards are the one case where you should stop asking for a building and start asking for a graphic.

Competition board collage, flat graphic style, of a small civic
pavilion in a park, single storey, mature trees on three sides,
containing a cafe and public toilets, a shallow folded-plate roof
over a glazed box, exposed glulam and glass with a copper fascia,
bright even daylight, cut-out scale figures in flat colour,
straight-on elevation with a thin section cut alongside,
generous white space, no gradients, no drop shadows,
--ar 16:9 --s 400

If the board needs legible labels, this is where model choice starts to matter. Google's documentation for Gemini 3 Pro Image, the model marketed as Nano Banana Pro, positions it for "complex graphic design, high-fidelity product mockups, and factual data visualizations that require accurate text rendering", and lists 1K, 2K and 4K output (ai.google.dev, checked August 26, 2026). Midjourney does not make an equivalent claim about text.

Also expect to redo the text by hand. Every current model garbles some proportion of small type, and a board with a misspelled room name is worse than a board with no labels at all.

Which model should you point the generator at?

The block is model-agnostic. Only the tail changes. Here is what each vendor documents, checked at their own pages on August 26, 2026.

Practical reading of that table for architects: Midjourney for look development and anything where you want to lock a house style across a set. Nano Banana Pro when the image carries text or you need 4K for print. GPT Image 2 when you have an image already and want to change one part of it, since inpainting is listed among its supported features on OpenAI's model page.

The tails, ready to paste:

# Midjourney: parameters go at the very end, after a space, with no
# punctuation inside the parameter itself
<assembled sentence> --ar 3:2 --s 250 --raw --no cars

# Gemini / Nano Banana Pro: plain prose, then state ratio and size
<assembled sentence>
Aspect ratio 16:9. Resolution 4K.

# GPT Image 2: plain prose; to change one element, use the edits
# endpoint with a mask rather than rewriting the whole prompt
<assembled sentence>

One caveat on the Midjourney figures. The parameter list article that documents --hd and --sd was last updated on June 11, 2026 and still describes them in terms of V8.1, while the version article confirms V8.2 became the default on July 24, 2026. The two pages disagree about which version those flags belong to. Check the parameter list yourself before quoting a resolution to a client. If your Midjourney output keeps drifting from the prompt regardless of parameters, the common failure modes are catalogued here.

What this generator cannot do

This is the section that matters more than the prompts, and the reason we would rather you bookmarked this page than a prompt list.

Image models generate pixels. They have no model of geometry, no dimensions, no awareness of your jurisdiction's building code, and no concept of load. Specifically, they cannot give you:

• Floor plans you can use. A generated plan is a drawing-shaped image. Rooms will not tile, circulation will not connect, and dimensions are decorative.
• Code compliance. Egress travel distances, fire compartmentation, accessible clearances, stair geometry, setbacks and plot ratio are all invisible to the model.
• Structure. Columns land where they look good. Spans are whatever the composition wanted. Cantilevers are free.
• A consistent building. Two views of "the same" scheme are two different buildings that resemble each other.
• Site truth. It has never seen your site, and it will confidently supply a neighbouring context that does not exist. This is ordinary hallucination, applied to your street.

If you need software that outputs editable geometry rather than pictures, that is a different category and those tools exist. Checked at their own sites on August 26, 2026: TestFit bills itself as a real estate feasibility platform whose Site Solver generates site plans "you can edit down to the last parking stall"; Autodesk's Forma Site Design describes itself as "AI-powered cloud software for site planning and analysis"; and Veras, from EvolveLAB, is "an AI-powered visualization app that plugs into your design authoring app", running on Revit, SketchUp, Rhino, Forma, Archicad and Vectorworks, and using your existing 3D model as the substrate. That last one is the interesting bridge, because it renders your geometry instead of inventing new geometry.

We do not compete with any of them, and we are not going to pretend otherwise. Prompt Architects makes prompts better. It does not make images.

How do you stop rebuilding this block every time?

Once the block works, the failure mode is administrative rather than creative. It ends up in a Slack message, then a notes app, then nowhere, and three weeks later you rewrite it from memory and it is slightly worse.

Three things fix that, in increasing order of effort:

1. Freeze the block, vary the slots. Keep one canonical version. Never edit it in place while working on a live project. Copy, fill, use.
2. Turn the recurring slots into named variables. Practice name, city, house drawing style, default aspect ratio and the two or three materials your office actually specifies do not change between projects. They should not be retyped between projects either.
3. Store it where your team can find it. A shared library beats a personal one, because the reason prompts decay is that the person who wrote the good version left.

That third point is not theoretical for us. In our own analysis of 2,170 customers on July 15, 2026, the average customer used 1.16 of our 7 features, and library adoption dropped from 69.7% on the core enhancer to 23.8%. Most people never get to the storage step, and most people rewrite their prompts forever as a result. If you want the general version of this, we wrote how to build a personal AI prompt library.

Prompt Architects is where we would obviously point you for steps two and three: save the block once as a prompt template, put your practice details into global variables, and share the result with your team. The free plan includes 5 prompt enhancements per day, forever, per our pricing FAQ, and paid plans start at $4.99 per month at the time of writing. Check current pricing before you decide, because the launch discount is temporary.

What we will not tell you is that any of this designs the building. The eight slots are a way of writing down decisions you have already made, quickly enough that testing a variation costs a minute instead of an afternoon. The judgement is still yours, and so is the drawing set.