Project Microsite · v4

Version diff

Every version of this project, what changed between them, and why. Each version page is preserved and linked.

The diff

Version	Focus	What changed vs prior	Link
v1	First build	Standalone microsite: spec, tool catalogue, reference imagery, placeholder slots. Standalone dark theme.	v1 →
v2	Renders + 3D + repair	House style; real airdropped photos replace references; Nano Banana clean renders; 5 image/text-to-3D routes tested; Route B mesh repaired to watertight (15 errors → 0); evolution narrative; version links.	v2 →
v3	Parametric CAD + colour	Adds text-to-CAD AI (Zoo, AdamCAD, CADGPT, Text-to-CadQuery, Claude, Fusion) — the editable/dimensioned route mesh AI can't give; two colourways from locked prompt v1.0.	v3 →
v4	Version diff	This page — captures the full evolution and links every version.	—

What changed, by version

v1 → v2

The biggest jump. v1 was assumptions and stock references; v2 became real once you airdropped photos. Generative text-to-3D failed loudly (0/10 blob), photo-driven image-to-3D worked, and the "15 errors" mesh was repaired to a watertight model. Doors, wheels (outboard, 265 fenders), axle drop and construction (steel frame + 4 mm ACP) all corrected.

v2 → v3

Closed the one gap mesh AI can't fill — exact, editable geometry — by adding the text-to-CAD class, and committed the two colourways (RAV4 blue, glossy red) under one locked prompt so details stop drifting.

v3 → v4

Documentation: this diff, so the decision trail is legible and every version stays reachable.

Locked fact sheet

Confirmed (2026-06-22). Body 1500 × 1500 × 2000; 2500 overall; flat front face 430, centred, full height. Steel frame + 4 mm ACP, bonded. Doors: 2 rear barn (900 + 600, off-centre) + 1 full-height on the RH tapered prow face; black RV paddle handles, no external hinges; long sides blank. Wheels 265/70/16 outboard, 265 fenders, axle 800 from rear and 100 below base. 550 W roof solar, flat, centred. All trim shiny polished aluminium. Colourways: RAV4 steel-blue · glossy red.

Open: belt-line confirmation (height/extent). Source: prompt v1.0.

Lessons captured

Match the tool to the goal. A picture (Nano Banana), an editable mesh (image-to-3D), and a dimensioned CAD model (text-to-CAD / Fusion) are three different jobs.
Feed real input. Generative AI reads intent, not millimetres — photos of the real object beat any text prompt for geometry.
Lock the prompt. Rewriting it freehand each run made details drift (an invented door). One versioned prompt, edited deliberately.
Mesh errors are correctable. "15 errors" was mostly unmerged duplicate faces — repair + remesh took it to watertight.