---

name: model-prompt-pack description: "Use when asked to write prompts for image-to-3D or text-to-3D tools, generate 3D models of a topic's parts, plan the assets for a 3D explorer, or build a labelled parts list with sources. Produces prompts for image-to-3D tools, one per object, plus a parts list for labelling in which every label names a trusted source, and states that generated geometry is illustrative, not authoritative. Second step of explorer-pipeline." version: 1.0.0

Model Prompt Pack

Image-to-3D tools can turn a clean reference image into a usable mesh in minutes, but they invent detail: an extra valve, a wrongly placed bone, a plausible part that does not exist. For an explorer that teaches, the geometry can be approximate; the labels cannot be wrong. This skill writes one prompt per object for generating reference images and meshes, and a parts list in which every label is tied to a named trusted source.

The accuracy rule

1. Every label needs a named trusted source: a textbook with edition, a standards body, a peer-reviewed paper, a museum or manufacturer reference, or an educational body's published material. A URL alone is not a source unless it identifies the publisher.
2. Generated geometry is illustrative, not authoritative. Every output that shows the model says so, and the viewer displays the statement.
3. No source, no label. A part the user cannot source is either dropped or shown as "unlabelled".

Required Inputs

Ask for these if not provided:

• Topic and object(s): one assembly, or several separate objects
• Level of detail: which parts a learner must identify, from the interface brief if there is one
• Trusted sources available to the user, or the curriculum or standard to follow
• The 3D tool, if chosen (for example an image-to-3D service, or an open model such as TRELLIS or Hunyuan3D), and any polygon or file-size limits
• Style: realistic, stylised or diagrammatic

Output Structure

1. Asset plan

Decide, and say why: one mesh per part (generate each part separately, then assemble; best for exploded views and clicking) or one mesh split afterwards (generate the whole, then separate parts in a 3D editor). Most explorers need one mesh per selectable part.

2. Prompt cards, one per object

Object: [part name]                     Mesh name in GLB: [kebab_or_PascalName]
Reference image prompt:
  "[part], isolated, centred, three-quarter view, neutral light grey background, soft even
   studio lighting, [style], clear silhouette, no text, no labels, no shadow on background"
Negative prompt: text, watermark, multiple objects, cropped, motion blur, extra parts
Image-to-3D settings: single object; target about [n] triangles; PBR texture [size]; GLB output
Fit check: compare the generated mesh against [named source figure] for shape and position

3. Parts list for labelling

| Mesh name | Label | One-line description (learner level) | Source (publisher, title, edition or date, page or figure) | Confidence (high, medium) |

The mesh names in this table must match the names in the GLB exactly; explorer-viewer-builder reads them.

4. Assembly and export notes

Scale (one unit equals one centimetre, or state otherwise), origin at the assembly's centre, each part as a separately named node, total GLB size target (under 15 MB for the web), and Draco or meshopt compression if the viewer supports it.

5. Illustrative statement

The exact sentence for the viewer, for example: "This 3D model is illustrative. Shapes and proportions are approximate; labels follow [source]."

Quality Checks

• There is exactly one prompt card per object or part to be generated
• Every label in the parts list names a trusted source with publisher and title
• No label appears without a source; unsourced parts are dropped or marked unlabelled
• Mesh names in the parts list match the names the prompt cards assign
• Every prompt card includes a fit check against a named source figure
• The illustrative statement is included and names the source used for labels
• Export notes state scale, origin and a file-size target

Anti-Patterns

• Trusting the mesh. Generators add plausible detail; check each part against the source figure.
• Wikipedia as the only source. Use it to find the underlying reference, then cite that.
• Prompting for the labelled diagram. Text in generated images is unreliable and copies textbook layouts; generate clean objects and label in the viewer.
• One giant mesh. Without separate parts there is nothing to click or explode.

Example Trigger Phrases

• "Write image-to-3D prompts for the parts of the human heart, with a sourced parts list."
• "Plan the 3D assets for a jet engine explorer."
• "Give me a prompt per part for generating a plant cell in 3D."
• "I need labelled 3D models for a volcano cross-section. Sources must be cited."