One sentence in. A spec, a sizzle, and a quotable STEP file out.

The brief was a single line: “a stress toy that looks like a trillion-cut gem, translucent shell, blue glitter liquid, fits a 60 mm cube.” Everything below was generated, checked, and corrected by a pipeline in one evening — including the part where it caught its own 4× spec error.

The look, and the failure that taught us something

01 · Renders

The first render round failed the same way three times: the model heard “gem” and produced rigid crystal. The fix was not more tries — it was prompt order: describe the soft-body physics first and let the gem shape follow. That lesson is now a reusable rule in the pipeline's prompt library, and an automatic judge scores every render against the spec so a human never has to catch this class of failure again.

First-round render: reads as rigid cut crystal — Round 1 · rejected Reads as solid glass. The judge fails it: no soft-material cues.

Hero render with visible fill line and settled glitter — Round 2 · hero **Proves the fill.** Meniscus and settled glitter read as liquid in a shell.

Hand squeezing the toy, fingertips denting the shell — Round 2 · squeeze **Proves the squish.** Fingertips dent the shell; it bulges between them.

The sizzle

02 · Motion

Six seconds of squeeze, glitter storm, and slow rebound — generated from the locked hero render with a motion-only prompt. This is the pitch-currency asset: the thing a licensor reacts to before any tooling money is spent.

Seedance 2.0 · image-to-video Prompted with motion, not appearance: “fingers tighten … glitter erupts … the shell rebounds slowly.”

The catch: geometry falsified the spec

03 · Verification

The first spec sheet estimated the fill at 65–75 cm³ and the weight at 80–100 g. When the parametric CAD was built, the real cavity measured 15.8 cm³ — the estimate was wrong by 4×. That class of error normally surfaces after a mold quote. Here it surfaced the same evening, for free, and the spec was corrected to Rev B before anyone saw it.

The rule this produced Any number derivable from geometry is derived, never estimated. In the live pipeline, CAD massing runs before the spec sheet is published — the numbers on the sheet come from solids, not guesses.

Metric	Hand estimate (Rev A)	CAD-measured (Rev B)
Cavity volume	65–75 cm³	15.8 cm³ @ 2.5 mm wall
Fluid @ 93% fill	—	14.7 cm³
Finished weight	80–100 g	≈31 g
Envelope	60 mm cube (assumed tight)	51.1 × 45.4 × 42.0 mm

The geometry

04 · CAD

Parametric solids, not concept meshes: shelled wall at 2.5 and 3.0 mm brackets, girdle split with an ultrasonic-weld tongue and groove, a 3 mm fill port. The STEP files open in any CAD package a mold maker uses. Drag to orbit; the explode slider pulls the halves apart at the weld line.

Explode

What the live pipeline does

05 · Pipeline

The case study above was driven conversationally. The live version runs the same stages unattended, with the lessons baked in:

Compute-first spec

Claude turns the brief into a construction template and materials spec; every volume and weight is derived from parametric geometry before the sheet exists.

Renders with an automatic judge

Three shots, each scored by a vision judge against the spec's own criteria. A failed shot regenerates once with the judge's corrective prompt.

Manufacturing-grade CAD

Shelled solids, weld joint, fill port. STEP + STL exports, and a cross-check that the built geometry agrees with the published spec numbers.

Reference-locked sizzle

The judged hero render seeds the video, so the toy in motion is the toy in the spec. Sampled frames get their own judge pass.

Honest limits Generated CAD is concept-grade: quotable geometry, but a mold maker's DFM pass still owns draft and gating, and safety compliance stays human. The pipeline compresses the paperwork and the visualization — not the play.