Geometry Preview

Geometry Preview

Overview

Use this skill to turn geometry intent into cheap, inspectable artifacts before escalating to CAD kernels or solvers. Treat the workflow as guardrails, not a fixed recipe: prefer the fastest artifact that can answer the current uncertainty, and escalate immediately when the uncertainty is solver-specific.

Exploration Ladder

Work from light to heavy, but skip upward when the task already demands a heavier truth source.

1. Extract dimensions, units, parameters, coordinate frames, and topology invariants from the request.
2. Choose the smallest geometry representation that naturally expresses those invariants: analytic curve/surface, B-rep solid, semantic spline, graph-to-geometry generator, mesh, level set, or solver-native feature.
3. Generate a lightweight candidate with Python or browser-native geometry: build123d, cadquery, trimesh, pyvista, gmsh, meshio, or Three.js depending on what is installed and the artifact needed.
4. Produce review artifacts: at least one preview image or HTML viewer, plus an exported geometry artifact when useful (.stl, .step, .msh, .vtk, .vtu) and a short QA summary.
5. Check invariants directly: bounds, units, volume, surface/solid count, connected components, holes, contacts, gaps, clearances, self-intersections, non-manifold edges, repeated-cell continuity, and feature size versus target mesh size.
6. Escalate to a CAD kernel or solver only when the current uncertainty requires it: B-rep validity, STEP import, Boolean acceptance, meshing, named selections, physics/material/boundary binding, or solver-specific feature tags.

Topology First Gate

Before making a pretty render, separate source authority:

• numeric authority: dimensions, coordinates, material thickness, pitch, clearance, feature size
• topology authority: CAD, reference code, equations, physical sketches, known connectivity, manufacturing constraints
• visual-only reference: screenshots, PPT drawings, non-dimensional meshes, icons, arrows, captions
• calibration target: points or images used to fit parameters, not necessarily ordered path constraints

Then write the non-negotiable topology:

• continuity: one body, many bodies, repeated cells, closed surface, open shell
• inside/outside: cavities, holes, membranes, channels, porous regions
• adjacency: contacts, gaps, bonded faces, shared boundaries, clearance targets
• ordering: over/under paths, front/back depth, layering, sweep direction
• periodicity: cell pitch, boundary continuation, mirrored or rotated repeats

Do not connect arbitrary points just because they are labeled. Treat points as calibration targets unless the source explicitly says they are ordered path points. Treat screenshots and hand sketches as visual references unless they include explicit dimensions or coordinate data.

Review Views

Use only the views needed to answer the current topology risk. For fragile topology, generate more than one view before accepting the model:

• labeled single primitive or minimal repeat unit
• tiled or repeated cell when periodicity matters
• front view plus side/depth view when ordering or clearance matters
• back view or depth-colored view when front/back topology could be ambiguous
• exported geometry preview when CAD or solver transfer is expected

Tool Choices

Prefer pip-ready tools for the first pass when they fit the task:

• build123d or cadquery: parametric B-rep CAD, Booleans, STEP/STL export.
• trimesh: mesh creation, repair attempts, mass properties, STL/OBJ/PLY IO, bounds, component counts, and quick surface checks.
• pyvista: 3D visualization, screenshots, mesh inspection, VTK/VTU outputs.
• gmsh: geometry/mesh preflight, .msh generation, feature-size and meshing failures before solver import.
• meshio: mesh conversion between .msh, .vtk, .vtu, .xdmf, and other solver-adjacent formats.
• Three.js or another lightweight viewer: interactive HTML preview when a browser artifact is more useful than a static image.

Do not assume any tool is installed. Probe the project environment first and use the smallest install or fallback that fits the user's constraints.

Do not force this stack when another representation is more natural. A formula, existing CAD file, solver-native feature tree, or project-specific generator can be the better kernel.

Artifact Contract

For geometry-generation tasks, try to leave a small artifact set:

• source script or notebook cell that regenerates the geometry
• preview image or HTML viewer
• export file when useful: .step for CAD/COMSOL, .stl for surface preview, .msh for mesh preflight, .vtk/.vtu for visualization
• QA notes listing dimensions, units, topology invariants, checks run, and known limitations

Say clearly what each artifact proves. A mesh preview does not prove COMSOL can build the geometry; a STEP export does not prove physics selections are correct; a successful Gmsh mesh does not prove the solver mesh will be acceptable.

Keep the artifact set proportional. A simple bracket may only need a script, one preview, and a STEP file. A porous, interlaced, or repeated topology may need multiple views and invariant notes.

Solver Handoff Gate

Before calling a mesh solver-ready, verify the handoff rather than only the picture or mesher exit code:

• units, scale, bounds, dimensionality, region count, and connected components
• required physical groups or named zones exist, are non-empty, and have the expected entity dimension
• boundary orientation and inside/outside assignment are consistent
• periodic partners have compatible topology, ordering, orientation, and transform; visual similarity alone is insufficient
• quality evidence names the metric, worst value, and location; do not apply a universal threshold without the target solver and element type
• a solver import or native mesh-check dry run can inventory the same regions and boundary names without silently dropping entities

Report the strongest state actually proven: preview-ready, mesh-generated, solver-imported, or setup-ready. These states are not interchangeable.

Solver Escalation

Use heavy solvers as authority layers, not as the first sketchpad.

• Escalate to COMSOL when the user needs .mph creation or inspection, geometry build, selections, mesh, material/physics/boundary binding, Java API tags/properties, or solver results.
• Escalate to Fusion, FreeCAD, or another CAD kernel when B-rep validity, feature history, editable solids, or STEP round-tripping is the question.
• Escalate to Fluent, Mechanical, Workbench, Abaqus, or STAR-CCM+ when the geometry must be imported into that solver's workflow or checked against its meshing/setup constraints.

If the user explicitly asks to work inside a solver, or provides a solver-native model file, use the relevant solver skill directly and treat this skill as a preflight companion.

Failure Handling

• If a lightweight render looks right but topology checks fail, revise the geometry kernel instead of polishing the render.
• If B-rep export fails, reduce to the smallest failing primitive and record the failed operation.
• If continuity is only represented by captions, labels, arrows, or metadata, encode it in geometry or report that the model is only schematic.
• If a connector is patching a broken local primitive, revisit the kernel or parameterization instead of adding more connector patches.
• If solver import fails after lightweight checks pass, preserve both artifact sets and report the boundary: preview passed, solver-specific validation did not.
• If a task depends on software or licenses not visible in the environment, state the missing dependency and offer a lighter artifact that can still be produced.