Architecture Prompts for Gemini (What Changes vs ChatGPT)

TL;DR: Gemini's real edge for architects isn't a smarter chatbot. It's multimodal input. Photograph a site, scan a plan, paste a mood board, and prompt against what the model actually sees instead of describing it in words first. As of August 2026, Gemini 3.1 Pro handles that reasoning; Gemini can't produce accurate floor plans, code compliance, or structural work.

What Is an "Architecture Prompt for Gemini," and How Is It Different?

An architecture prompt for Gemini is an instruction paired with an image, photo, or scanned document, not a text description alone. That pairing is the entire point of using Gemini instead of a text-only chat window.

To be clear about the word first: this article is for building architects, not software architects or "prompt architecture" as a technical pattern. If you searched for the latter, you want a different article.

The shift that matters is prompting shape. A text-only prompt asks a model to imagine a site from your description of it: the slope, the neighboring rooflines, the light. A multimodal prompt hands the model a photo of the actual slope and asks it to reason about what's in the frame. Gemini was built natively multimodal from the ground up, and Google's own developer documentation describes the current Gemini 3 family as accepting "text, images, video, audio and code" as native inputs to a single model, not bolted-on separately (ai.google.dev, accessed August 2026).

ChatGPT reads images too. That's not new, and this article won't pretend otherwise. What differs is the workflow: how many files you can drop in one prompt, how the model treats a multi-page PDF, and how tightly reading an image connects to editing one. That's the comparison in the table below, not a "Gemini can, ChatGPT can't" claim.

What Should You Feed Gemini Before You Type Anything?

Feed it the actual visual material first, then write a short, specific instruction, in that order. The image or scan is doing most of the work; the text is just telling Gemini what to do with it.

Four inputs cover almost every early-stage architecture task:

• Site photos: taken on a phone, from the actual angles a design decision depends on, the approach, the worst-case shadow, the neighboring context you're deferring to or breaking from.
• Precedent images: a building, a detail, a landscape you're referencing, so Gemini can name what it's actually looking at rather than paraphrase a style label.
• Plan or elevation scans: hand sketches, trace paper, or a PDF export, for narrative and communication purposes, never as a stand-in for a checked drawing.
• Mood boards: a collage of material, color, and texture references, which Gemini can decompose into a written palette description faster than you can write one from scratch.

The phrasing habit that makes this work: ask Gemini to describe or reason about specific things it can see in the image: proportions, materials, light direction, adjacency, rather than asking a generic question the image doesn't actually inform. "What style is this house?" doesn't need the photo. "Given the roof pitch and eave depth in this photo, what does that suggest about the original construction era and climate response?" does.

The Prompt Architects browser extension supports Gemini natively alongside ChatGPT and Claude, so the same structured-prompt habit (role, task, constraints, tone) carries over to whichever tool has the file open, without rewriting your process per platform.

Site & Context Photos

Here's a photo of the building site, taken from the street at 4pm.
Based on what you can see — slope direction, existing tree cover, and
the neighboring rooflines — list three site conditions I should design
around, and explain the reasoning for each in under 40 words.

I'm attaching two photos of the same lot: one from 9am and one from 4pm
in the same season. Compare the shadow patterns between them and tell
me where on the site full sun is most reliable through the working day.

This is a street-level photo of the block my project sits on. Looking
at the massing, material, and roofline of the three adjacent buildings,
describe the context language a fourth building on this block would
need to either match or deliberately depart from.

Precedent & Concept Development

Here's a precedent image I'm referencing and a photo of my own site.
Looking at both, tell me which specific elements of the precedent —
proportion, material, roof form, threshold condition — would actually
translate to my site's conditions, and which wouldn't.

I'm attaching a photo of a building whose facade rhythm I like. Describe
the underlying proportioning logic you can see — bay width relative to
height, void-to-solid ratio, datum lines — in language I could hand to
a facade consultant.

Here are four images of buildings I'm drawn to for this brief. Looking
at all four together, identify the one or two design principles they
share, stated abstractly enough that they could apply to a different
material and climate.

Material and Palette Work

This is my mood board for the project — six images collaged together.
Write a one-paragraph material and color palette description a
contractor could read and understand, naming the specific materials
you can identify in the images.

I'm attaching two photos of facade material samples we're deciding
between for the client. Compare them on warmth, texture, and how each
would read from 30 feet away, and flag anything that won't photograph
well for marketing.

Here's a close-up photo of a textured concrete panel. Describe its
tactile and visual qualities in the kind of language I'd use in a
specification narrative, not a marketing brochure.

Plan and Diagram Narratives

This is a scan of my hand-drawn bubble diagram for the ground floor.
Describe the circulation logic and adjacency relationships you can see,
and note any adjacency that looks like it creates a dead-end or
bottleneck — as a discussion point, not a fix.

Here's a rough parti sketch I drew for the client meeting tomorrow.
Turn what you see into two sentences of plain-English design narrative
I can say out loud, without inventing dimensions or room counts that
aren't in the sketch.

I'm attaching a concept sketch and a site photo side by side. Looking at
both, does the sketch's orientation and massing look coherent with the
site conditions in the photo, or is there a conflict I should flag before
we develop this further?

Client-Facing Presentation Language

Here are three render images from our competition board. Write the
design narrative paragraph that would sit beside them, describing the
spatial sequence and material story you can see across the three images,
in a confident but not overwritten tone.

I'm attaching our mood board and an early concept sketch together.
Combine what's in both images into a single one-page concept statement
a client with no architecture background could read in under a minute.

Here's a rendering we're about to present. Explain the design decision
behind the roof form shown, in the plain language I'd use with a client
who has already said they find angled rooflines confusing.

Multi-Image Comparison and Research

Here are two precedent photos from different countries and eras.
Looking at both, extract the one design principle they share that
isn't about style — something about how each handles light, threshold,
or scale — and name it in a sentence.

I'm attaching two photos of the same courtyard: one from spring and one
from winter. Write the landscape narrative paragraph for our submission
that accounts for how the space reads differently across both seasons.

Here's my elevation sketch and a precedent photo I'm referencing for
proportion. Looking at both together, write a short brief for the
facade consultant describing the proportion and material relationship
I'm going for, without stating exact dimensions that aren't in the sketch.

Gemini vs ChatGPT for Architecture Prompts: Same Task, Different Shape

Both tools read images today, so the honest comparison isn't access. It's ergonomics: how many files fit in one prompt, how each treats a scanned document, and how tightly reading connects to editing.

What Gemini Still Can't Do for Architects

Gemini cannot verify code compliance, run a structural check, or hand you an accurate, dimensioned floor plan. None of the multimodal input changes that. It reasons about what it sees; it doesn't calculate loads, check setbacks against a zoning ordinance, or guarantee that a room in a generated image is actually the size it looks like.

The image-editing side (Nano Banana Pro) can restyle a rendering or blend a mood board into a concept image, which is genuinely useful for early client conversations. It is not a CAD tool, and nothing it outputs should be scaled, measured, or built from directly.

Use Gemini the way this article's prompts do: to read what's already in front of you and turn it into language, comparison, or a first-draft narrative. Always keep a licensed professional reviewing anything that touches a real drawing set.

Where to Go Next

This post is deliberately narrow. It's the Gemini-specific layer on top of a broader practice. For the fuller picture of how architects are using AI prompting day to day, across tools, read AI Prompts for Architects.

For the image-generation side of this, actually producing a rendered concept image rather than reasoning about an uploaded one, Midjourney v7 Prompt Examples and Midjourney Style Modifiers cover the parameter and modifier vocabulary that carries across most image models, Gemini's included in spirit if not in exact syntax.

If you're starting from a precedent image rather than a site photo, How to Reverse-Engineer Any AI Image into a Reusable Prompt walks through the extraction habit this article applies to architecture specifically. And if your team wants the output structured rather than prose (for handoff to a script or a spec template), JSON Prompts Explained covers the format Gemini and most other models handle natively.