Our work 02 / 03 Architecture · Computational design

AI as a Modeling Assistant

In Rhino, by prompt, under constraint. An investigation into how an architect orchestrates AI inside their existing software. Geometry stays the source of truth.

Project shape
100s
Constraint-aware decisions per building
5principles
Methodology, codified
1:500scale
Watertight mesh, FDM print
MIT
Every prompt, file, output open
The premise

Buildings are not one prompt.

Large language models promise design-by-prompt. Buildings don't work that way. They take shape through hundreds of constraint-aware modeling decisions, not a single instruction.

This investigation reframes the role. AI is not the designer; it is a modeling assistant that operates inside the architect's existing software, executes single-objective instructions, and stays verifiable at every step.

The work gets denser, not easier. Where the assistant fails on multi-objective tasks, the architect's judgment is the limiting factor. AI does not make that judgment cheaper. It makes it the only thing left.

The methodology

Five principles for keeping AI verifiable.

Principle 01

One verb, one shape, one rule.

Single-objective prompting. Multi-objective prompts produce noise. Every instruction names a single modeling action: draw, divide, fillet, offset, boolean, mesh, export.

Single-objective prompting
Principle 02

Declare the constraints before the task.

Hard rules. Height, footprint, clearances, units, layers. Are declared numerically before any modeling task. The constraint set is the contract the assistant works inside.

Constraint-first framing
Principle 03

Trust the viewport, not the assistant's account.

Every result is inspected in the Rhino viewport and checked against the constraint set. The assistant's natural-language summary is treated as unverified until the geometry confirms it.

Viewport verification
Principle 04

Branch before you commit.

Each iteration is duplicated to a new sublayer. The architect can return to the last good version whenever an experiment fails. Failure stays cheap, exploration stays bold.

Layered checkpointing
Principle 05

Failure is the map.

The point where the assistant breaks marks the boundary of its competence and the start of the architect's. Each failure mode enters a prompt library. The methodology compounds.

Failure as method
The pipeline

One source of truth, three downstream artefacts.

The geometry built under prompt is the canonical model. From it, everything downstream is a transformation, not an interpretation. What the architect approved in the model is what appears in the photograph and the printed artefact.

Geometry under prompt

The Rhino 8 model is the source of truth. Every modeling operation is logged, every iteration checkpointed to a sublayer.

Photoreal context

Image generation produces site-aware renders without moving any mass. The model the architect approved is the model that appears.

Watertight mesh

The same model exports a watertight mesh for desktop FDM 3D printing at architectural scales. 1:500 to 1:1000.

Prompt library

A documented workflow with failure-recovery variants. The method is transferable; each new project compounds the library.

The stack

Off-the-shelf parts, methodically composed.

Modeling Rhino 8 The architect's existing surface. No new tool.
Assistant Claude (LLM) via Rhino MCP plugin Connected live: reads scene state, calls modeling functions.
Image generation Gemini 2.5 Flash Context renders that preserve geometry. No mass moved.
Physical output Desktop FDM 3D printing Architectural model scales · 1:500 to 1:1000.
License MIT. Open from day one Every prompt, file, and output is public.
Applicable to

The methodology travels.

Constraint-first, single-objective, viewport-verified is a discipline, not a tool. It applies anywhere a domain expert needs an LLM to operate inside high-stakes software without losing audit trail.

Concept massing studies Urban constraint testing Component library construction Studio tooling for computational design LLM-driven CAD experiments Open-source design pipelines
"AI does not make the architect's judgment cheaper. It makes it the only thing left."
Practice note · AIHLPR · Architecture engagement
Next case study 03 / 03
Workforce · Owned product

CareerOS