Plate I · a method, not an app

Photographs in.
A reasoned Blender model out.

Specimender turns Claude Code into a 3D reconstruction operator. It studies your photos, writes an anatomical description of the object, resolves every doubt with you in plain language, then builds the model in Blender — one small, reviewable script per part.

No point clouds. No mesh soup. A specimen study you can open, understand and edit.

View on GitHub Read the method

Blender render of specimen 001, right lateral view

Fig. 1 — Specimen № 001 « pipeau », right lateral view. Reconstructed in Blender from seven phone photographs, without a human ever opening Blender.

II.

The record

from camera to specimen

Phone photograph of the real object on graph paper

Observed. Fig. 2

Seven ordinary phone photos. Each one is described in standard anatomical language — cranial, caudal, ventral, dorsal — before anything is built.

## 4. Réservoir

Corps de révolution d'axe craniocaudal, s'étendant de Z = 90 mm à Z = 130 mm. Embase tronconique métallique, se rétrécissant vers le caudal […]

## 5. Embout

Pièce tronconique noire s'évasant vers le cranial, surface externe facettée en nid d'abeille, motif hexagonal régulier […]

Specified. The contract

Descriptions are correlated into one anatomical specification. Every uncertainty becomes a question for you. The validated text is the contract.

Control plate: three anatomical views of the Blender model

Built & criticized. Fig. 3

One idempotent Blender script per part, driven through MCP. Control renders are compared against the photos by a critic agent until you call it done.

III.

Why anatomy?

the origin of the method

By the accidents of life, the author's only rigorous descriptive skill is anatomical description — not 3D modeling, not Blender. Specimender makes that constraint its founding principle: the human describes and validates the object the way a naturalist describes a specimen, and the AI translates that language into geometry.

It turns out anatomical terminology is an excellent specification language for man-made objects too. It is standardized, unambiguous about orientation, and precise about shape and relief — crests, grooves, fossae, fillets that flare toward the cranial end. It survives translation between human and machine with nothing lost.

And what is Specimender, then? Not an application: no binary, no interface. A documented method — a driver made of prompts, skills and a few small scripts — sitting between Blender (the software) and Claude (the intelligence).

« The human never touches Blender ; the AI never fills a gap in silence. »

IV.

The pipeline

18 steps, four movements

Phase	Steps	What happens
Observe	P0 – P4	Photos ingested and resized; each one described anatomically; views correlated, orientation fixed once and for all.
Specify	P5 – P11	Full anatomical description; uncertainty register; your answers; validated specification; dimensional calibration; part plan.
Build	P12	Part-by-part construction in Blender via MCP — small idempotent scripts, scene saved after every part.
Criticize	P13 – P18	Control renders vs photos, critic agent, iterations, delivery (GLB, STL, OBJ, FBX), lessons learned.

Against the point cloud

photogrammetry alternative

Automatic photogrammetry

Input — dozens of well-lit, overlapping shots
Output — millions of frozen triangles
Occluded zones — guessed, or holes
Editability — sculpt-repair at best
Paper trail — none

Reasoned reconstruction

Input — a handful of phone photos
Output — separate parts, true curves, fillets, symmetry
Occluded zones — resolved by asking you
Editability — every part is a reviewable script
Paper trail — descriptions, doubts, dimensions, decisions

Photographs in.A reasoned Blender model out.