Render Prompt

Render Prompt

Craft an optimised architectural render prompt from a design description.

When to use

When an architect, designer, or homeowner needs an AI image render and wants the prompt to actually produce what they intend. This skill translates loose design intent into a precise, structured prompt that works across image models. It crafts the prompt; it does not generate the image.

The single biggest quality lever is material specificity — a named material and finish beats any adjective. The second is describing light behaviour (raking, diffused, grazing) rather than mood words. Structure and negatives matter, but those two dominate.

Input

The user provides a design description. This can be:

• A single sentence: "minimalist concrete house, golden hour"
• A detailed brief: room type, style, materials, mood, time of day
• A reference to a style movement: "Tadao Ando meets Scandinavian warmth"
• An existing render they want to iterate on

Process

1. Parse the intent. Identify: building/room type, style, dominant materials, lighting condition, camera perspective, mood.

1. Fill gaps with informed defaults. If the user says "modern kitchen" but doesn't specify lighting, default to soft north-facing daylight — the most flattering for kitchens. Always choose the option that an experienced architectural photographer would choose.

1. Structure the prompt using this hierarchy (most to least important for image models):

• Subject: what is being rendered (room type, building type, scale)
• Style: architectural movement or aesthetic direction
• Materials: specific materials, not generic ("white-washed Douglas fir ceiling" not "wood ceiling")
• Lighting: time of day, direction, quality (diffused, raking, golden)
• Camera: eye level, bird's eye, worm's eye, focal length equivalent
• Atmosphere: mood, season, weather, human presence
• Technical: resolution, aspect ratio, rendering style (photorealistic, watercolour, section)

1. Output the prompt in two formats:

• Full prompt: detailed, comma-separated, ready to paste
• Negative prompt: what to exclude (distortion, oversaturation, unrealistic scale)

Domain knowledge

Lighting conditions by space type

• Living rooms: late afternoon west light, warm, long shadows. Or: overcast north light for Scandinavian interiors.
• Kitchens: morning east light or soft north light. Avoid direct overhead — it kills material texture.
• Bathrooms: soft diffused light, steam atmosphere. Dawn or dusk.
• Exteriors: golden hour (low sun, warm) or blue hour (dusk, moody). Midday flattens facades.
• Bedrooms: early morning, curtains filtering light. Intimate, not staged.

Camera perspectives

• Eye level (1.2m): most natural, residential feel. Default for interiors.
• Seated level (0.9m): living rooms, restaurants. Creates intimacy.
• Elevated (2.5m): shows spatial layout. Good for open plans.
• Worm's eye (0.3m): dramatic, emphasises ceiling and volume. Use for double-height spaces.
• Bird's eye: aerial exterior views. Requires explicit request.

Material specificity

Never use generic material names. Translate:

• "wood" → "white oak with matte finish" or "charred Yakisugi cedar" or "reclaimed Douglas fir"
• "stone" → "honed Ceppo di Gré limestone" or "bush-hammered Vals quartzite"
• "concrete" → "board-formed concrete with vertical grain" or "polished microcement in warm grey"
• "metal" → "patinated brass hardware" or "blackened steel window frames" or "brushed stainless"
• "tile" → "Moroccan zellige in sea foam" or "matte white subway with dark grout"

Style shorthands

When the user references a style, expand it:

• Japandi: low furniture, natural materials (oak, linen, stone), muted earth palette, negative space, handmade ceramics
• Brutalist: raw concrete, dramatic volumes, angular geometry, raking light, minimal furniture
• Warm Minimalism: travertine, bouclé, matte plaster, coffee/mocha palette, soft curves
• Scandinavian Modern: light wood (ash, birch), white walls, wool textiles, functional furniture, north light
• Mediterranean Modern: lime wash walls, terracotta, olive wood, linen, outdoor-indoor continuity
• Mid-Century Modern: walnut, brass, terrazzo, statement lighting, low-slung furniture, warm tones

Holding the same building across views

Image models keep no 3D model between generations, so "the same building from the south" tends to return a new building. To hold a design across views:

• Reference image / ControlNet is the only reliable geometry lock. Condition on a massing screenshot, clay render, depth map, or line drawing so the model restyles without moving walls. This is the difference between concept mood and iterating a specific scheme.
• Seed holds broad composition and volume across small prompt changes on the same model — good for swapping materials or light on one view.
• Style reference (a fixed reference image, or a curated 30–60 image board of the practice's own work) holds palette and atmosphere so new views read in one visual language.
• Be honest about the ceiling: for true multi-view consistency of a specific building, you still need the 3D model. Text alone re-rolls.

Negative prompts: target real defects

Three to five sharp exclusions beat twenty vague ones. The discipline: generate first, look at what actually broke, then exclude that. A reliable base for architecture: distorted perspective, warped lines, impossible geometry, plastic-looking materials, oversaturated, blown highlights, people, cars, text, watermark. Don't stack fears you haven't seen — every weak negative dilutes the ones doing real work.

Model choice

Model strengths and names change fast, so they live in [references/model-notes.md](references/model-notes.md), not here — load it only if the user wants a tool recommendation, and treat it as a snapshot to sanity-check against today's tools, not gospel. What's durable is in this file: structure, material specificity, light behaviour, and the consistency techniques above.

Output format

RENDER PROMPT
─────────────
[Full structured prompt, 40-80 words, comma-separated]

NEGATIVE PROMPT
───────────────
[What to exclude, 10-20 words]

PARAMETERS
──────────
Aspect ratio: [e.g. 16:9, 3:2, 4:5 — always set one; unset ratios distort proportions]
Consistency: [reference-image/ControlNet to hold a specific design; seed to iterate one view]
Model family: [general concept model vs geometry-locked/plugin tool — see references/model-notes.md]

What you don't do

• Don't generate the image. You craft the prompt.
• Don't suggest styles the user didn't ask for unless filling a clear gap.
• Don't use "beautiful" or "stunning" — these are noise words in image prompts. Be specific.
• Don't output more than one prompt unless the user asks for variations.
• Don't imply a render is buildable. A prompt produces an image, not a design — geometry, code, and specification are not implied by a good picture.
• Flag the IP reality when it matters: under current US Copyright Office guidance (2025), a purely AI-generated image is not copyrightable without meaningful human authorship — relevant before an architect puts one on a competition board or in a proposal.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gaia-computer-technologies
• Source: gaia-computer-technologies/architecture-skills
• License: MIT
• Homepage: https://www.gaia.computer/skills

Install and usage instructions live in the source repository linked above.