Rhino Architectural Reverse Modeling

Rhino Architectural Reverse Modeling — Mode A

A [Claude Code](https://claude.ai/code) skill for turning reference photographs of a real building into a Rhino model that is **editable** and **dimensionally defensible**.

Not photogrammetry. Not a pixel-perfect copy. Not a single carved solid.

What it optimises for

Most image-to-model workflows optimise for *looks like the photo*. That is easy to reach and hard to build on: the result is usually one boolean-carved blob whose dimensions are eyeballed, and which nobody — including the model that made it — can tell you is wrong.

This skill optimises for two different things:

• Editable — one architectural decision equals one named, editable object. Changing the loggia depth must not require deleting the model and re-running a script.
• Dimensionally defensible — every load-bearing dimension is either measured two independent ways, given by the user, or explicitly flagged as unevidenced and asked about. Inventing a principal dimension and labelling it low-confidence is forbidden.

Design

It is a **constraint set, not a procedure.**

Earlier versions prescribed staged workflows (`LEVEL_1/2/3`, then `Phase A/B/C`). Staging turned out to be scaffolding for a weaker model that becomes drag on a stronger one — it spends attention on process that should go into looking at the building. It was also ordered by geometric scale, which does not correlate with where the risk is: the largest dimension in a building can be the one with no evidence at all.

So the staging was deleted. What remains is what does not lose value as the underlying model improves:

Kept	Why
Domain facts	typological dimension ranges, construction logic — things a model will not reliably recall
Executable checks	each one can fail; each caught a real error
Output constraints	editability, naming, no fabricated dimensions
Tool traps	silent failure modes in the Rhino MCP layer

Every exhortat