Create and review industrial design and physical product design from brief through concept selection, product architecture, proportion, form language, ergonomics, human factors, controls, interfaces, CMF, parametric CAD, refined native 3D, surface quality, product rendering, motion states, engineering validation, DFM, and revision-bound GB/T handoff. Use for consumer products, consumer electronics, appliances, smart-home hardware, enclosures, tools, equipment, rotating mechanisms, reference-image redesign, aesthetic direction, prototype planning, 3D product-quality review, or any task where an agent must design or improve a physical product before mechanical drafting and manufacturing release.
Create and review industrial design and physical product design from brief through concept selection, product architecture, proportion, form language, ergonomics, human factors, controls, interfaces, CMF, parametric CAD, refined native 3D, surface quality, product rendering, motion states, engineering validation, DFM, and revision-bound GB/T handoff. Use for consumer products, consumer electronics, appliances, smart-home hardware, enclosures, tools, equipment, rotating mechanisms, reference-image redesign, aesthetic direction, prototype planning, 3D product-quality review, or any task where an agent must design or improve a physical product before mechanical drafting and manufacturing release.
Industrial Product Design with GB/T Handoff
Treat product design as the primary task and mechanical drafting as the engineering expression of a sufficiently resolved design. Decide who the product serves, what it must do, how it is used, why it has its form, and how it will be made and verified before detailed CAD and drawings.
Core Principles
Converge function, architecture, interaction, form, CMF, manufacturing, service, and safety together. Do not apply styling as a late shell over an unresolved package.
Separate observation, inference, assumption, and confirmed requirement. Mark unknown inputs TBD and checks that did not run NOT_EVALUATED.
Create meaningful alternatives. Color or fillet variants of the first idea do not count as concept exploration.
Express aesthetics as testable form language: proportion, silhouette, mass hierarchy, geometry vocabulary, line and surface relationships, transitions, boundaries, detail density, CMF, and 360-degree coherence.
Keep editable parametric native CAD as engineering truth. Renders, meshes, images, plots, and exchange files are derived evidence.
Never infer manufacturing-critical dimensions, material performance, tolerances, safety distances, or compliance from appearance alone.
Lock the product's immutable design DNA before optimization. Do not solve usability, packaging, cost, or engineering pressure by silently deleting the defining innovation.
Label every output by maturity. A concept is not a manufacturable design, and automated checks do not grant release approval.
Never treat body count, entity count, dimension audit, geometry DRC, or save success as proof that proportion, form language, interface integration, motion logic, surface quality, or product realism is acceptable.
Portability Contract
Discover the available CAD application, automation bridge, renderers, parsers, and file formats before execution. Map capability-level steps to tools that actually exist.
Never invent a tool call or verification result. After mutable CAD operations, read back actual entities, parameters, associations, hierarchy, and geometry.
Preserve the user's source and unrelated geometry. Use a traceable working copy for major revisions; do not overwrite, remove, or replace user files without authority.
Keep deliverables and evidence in standard CAD, exchange, image, and report artifacts that another tool or reviewer can reopen.
Reference Routing
Before using web, standards, supplier, research, repository, or catalog evidence, read references/research-authority.md.
For every new design or substantial redesign, read references/design-brief-and-gates.md.
When a person uses, carries, installs, cleans, services, approaches, or works around the product, read references/human-factors-and-usability.md.
For form, product character, proportion, detail, or CMF, read references/form-language-and-aesthetics.md.
When the user supplies reference images, competitor images, sketches, or aesthetic feedback, read references/reference-image-learning.md and references/aesthetic-profile.md.
Before selecting CAD, surface, mesh, exchange, or rendering tools, read references/backend-routing.md.
For refined industrial-design 3D, ports, controls, rotating products, motion states, or presentation renders, read references/industrial-design-3d-quality.md.
For a large local CAD/JPG corpus, read references/local-cad-reference-library.md. Generate a private manifest with scripts/catalog_reference_library.py; use references/local-reference-library.example.json as the portable schema and do not commit personal absolute paths.
For concept selection, stage review, prototype review, or design audit, read references/integrated-design-review.md.
After user rejection, a gate escape, repeated misalignment, unequal repeated features, impractical layout, or an unattractive result, read references/failure-learning.md and update the owning gate.
When importing outcomes or feedback from another task/window, or deciding whether a lesson should persist across projects, read references/project-learning-loop.md.
For a new consumer product, enclosure, appliance, controls, ports, cables, moving parts, or mains-powered product, also read references/consumer-product-concept.md.
For parametric parts, surfaces, assemblies, motion, or 3D delivery, read references/cad-3d-modeling.md; for evidence tiers and 2D/3D workflows, read references/cad-workflows.md.
For complex mechanisms or multi-system assemblies, read references/complex-assembly-drafting.md; for physical-performance claims, read references/engineering-analysis.md.
For DRC, DFM, fit, interference, tolerance, or release review, read references/drc-review.md; for GPS, GD&T, or inspection, read references/gps-inspection.md.
Before GB/T engineering drawings, read references/gbt-drafting.md; for AutoCAD or AutoCAD MCP execution, also read references/autocad-mcp-workflow.md.
For configurations, BOMs, revisions, MBD/PMI, dependencies, or formal packages, read references/product-definition-release.md.
Load only the references required for the current task.
For Chinese-language drafting work, prefer the matching file under references/zh-CN/ when one exists; do not load both language versions unless comparing translations.
Upstream and Downstream Contract
Use this skill upstream to select product intent, architecture, proportion, interaction, form language, interfaces, motion states, CMF, and an approved 3D design revision.
Use mechanical-drafting-gbt downstream to complete manufacturing definition, GB/T drawings, dimensions, fits, GPS/GD&T, BOMs, DRC/DFM, exchange verification, and release evidence.
Do not enter manufacturing drawing work while the industrial-design gates for the claimed maturity remain failed or unevaluated.
Do not duplicate the complete GB/T rule set here; never weaken or bypass it at handoff.
CAD Reliability Gates
These gates apply before and during native 3D work. A successful command response never overrides them.
P0 document identity: After create, open, activate, save, export, or render, read back document ID, requested and active paths, revision, configuration, and units. Every mutation carries document ID and expected revision. Divergence returns E_DOCUMENT_ID_MISMATCH and stops the stage without silently switching a user document.
P0 atomic state: Treat the semantic feature registry, native entity registry, and B-rep as one transaction. A failed feature, stale revision, missing layer, backend exception, timeout, or postcondition mismatch rolls back geometry, registry rows, layer changes, temporary outputs, and journal pointers. Prove rollback with before/after counts and revision.
P0 postconditions: Compare requested and actual kind, parameters, associations, stable IDs, bounds, volume, layer, and configuration transform after every mutable operation. Any unexplained diff or unknown handle is a hard stop; permit one bounded recovery, then change backend or mark BLOCKED.
P1 interference semantics: Label AABB and sampled-motion results as broad-phase only. Separate containment, contact, permitted crossing, intentional motion overlay, and exact B-rep interference; geometry validity cannot substitute for an assembly or motion verdict.
P1 render truth: Bind each view to document, revision, configuration, fixed camera/projection, actual visual style, resolution, visible components, nonblank ratio, clipping state, content hash, and material_render_verified. Wireframe/linework output cannot pass a shaded or material review.
P1 prepared views: Section and exploded evidence require revision-bound section planes/cut sets or reversible component transforms. Ad-hoc camera movement is not a prepared section or exploded state.
P1 plot truth: FIT/NTS and fixed ratios are mutually exclusive. Verify paper, orientation, PDF MediaBox, viewport scale, and title-block declaration; any mixed claim is PLOT_SCALE_CONSISTENCY=FAIL.
P0 transport pressure: Route AutoCAD MCP calls through one FIFO lane with a bounded queue and sliding-window budget. Never launch parallel mutation or screenshot calls; when the guard returns a retryable limit/queue error, stop the round and preserve its evidence.
P1 response truth: Screenshot feedback returns a managed path, dimensions, byte count, and SHA-256 by default. Inline base64 requires explicit opt-in and a size ceiling; large entity/audit responses must be narrowed or read from saved artifacts instead of being replayed into the transcript.
Record title, publisher, edition or commit, date, retrieval, license or access limit, applicability, and bounded claim for decision-driving sources.
Treat official catalog pages as identity and scope evidence only when normative content is licensed or unavailable publicly.
Pin open-source CAD, geometry, exporter, and validator versions; rerun conformance fixtures after upgrades.
Route B-rep, surface, mesh, rendering, and drafting work to tools that can verify the required representation. Do not force one backend to impersonate missing capabilities.
Keep human factors, product-design review, engineering analysis, geometry DRC, format validation, and external approval as separate evidence channels.
Required Design Workflow
1. Classify the task and maturity
Classify the work as a new design, major redesign, local improvement, reconstruction, form study, structural design, mechanism design, engineering validation, or drafting handoff. Declare the target maturity:
discovery: user, scenario, need, and constraints are still being defined.
concept: product architecture, silhouette, and interaction directions are being compared.
design development: a selected direction is integrating structure, form, CMF, and process.
engineering candidate: critical interfaces, materials, processes, dimensions, and validation have evidence.
release candidate: applicable checks ran and unresolved items and external approval status are explicit.
2. Build the design brief
Record users, scenarios, task flow, functions, interfaces, loads and environment, anthropometric constraints, size, cost, process goals, service, safety, product character, family cues, retained features, prohibited features, and deliverables. Classify each requirement as:
supplied with authority;
derived from evidence;
assumed for exploration;
TBD;
contradictory and requiring a decision.
When inputs are missing, complete low-cost work that does not depend on them and constrain the maturity. Do not fill gaps with false precision.
Before detailed CAD, explicitly record user scenarios, interface types and counts, adapter or mating-object envelope assumptions, cable directions, motion actions and states, occupied footprint and context, tipping risks, product character, and immutable design DNA. Run scripts/check_pre_cad_brief.py when a structured brief is available. A failed PRE_CAD_DESIGN_GATE permits only research, sketches, provisional envelopes, and comparison massing.
3. Analyze references and existing product language
Use reference-image-learning.md. Record visible facts before interpreting intent. Add only user-confirmed preferences to the confirmed section of aesthetic-profile.md; keep agent inference labeled with evidence and confidence.
4. Establish system architecture and envelopes
Define functional blocks, energy, motion, signal and material flows, major components, purchased parts, interfaces, assembly direction, service paths, cables or hoses, thermal paths, motion sweeps, and hazard keep-outs. Create a named-parameter envelope skeleton before surface detail.
Create a dimension-authority register using levels A, B, C, and D from research-authority.md. Treat regulated or certification-critical interface geometry as confirmed only with level A controlled authority or level C controlled measurement. Public brand parameters and concept assumptions may guide exploration but cannot authorize detailed openings, terminals, protective devices, or moving electrical contacts. Run scripts/check_dimension_authority.py when the register is structured.
5. Generate and compare concepts
For a new design or major form change, create at least three materially different directions. Vary architecture, massing, center of gravity, posture, support, controls and interfaces, access, cables, or process strategy. Present each at comparable scale with:
one-sentence design intent;
user flow and critical actions;
primary masses, proportion, silhouette, and visual center;
internal architecture and critical envelopes;
benefits, tradeoffs, risks, unknowns, and likely processes;
links to the brief and aesthetic profile.
Treat hard constraints as pass or fail. Never hide a critical failure inside a weighted average.
6. Select a direction and define form grammar
Record selected, rejected, or combined concepts and rationale. If the user has not selected, deliver a comparison package and a recommendation. Continue along the recommendation only when selection authority is delegated or the task clearly permits provisional exploration; label the direction provisional.
Translate the direction into executable rules: primary, secondary, and tertiary masses; control lines; section strategy; curvature and radius families; seams and splits; openings and textures; interface integration; visual weight; CMF zones; and detail density. Check front, rear, sides, top, bottom, and expected use views.
7. Prove backend capability and document identity
Before refined 3D, route the task with backend-routing.md and pass BACKEND_CAPABILITY_GATE and DOCUMENT_IDENTITY_GATE from industrial-design-3d-quality.md. Require the active backend to support the operations and evidence needed by the claimed maturity. A primitive-only backend may create massing but may not claim refined product surfaces or product rendering.
After every open, create, switch, save, export, or render request, read back the active document ID, absolute path, revision, configuration, and output identity. If requested and actual identity differ, stop immediately, attempt one bounded recovery, and then switch to a verified backend or report blocked. Never audit or decorate the wrong document.
Run scripts/check_document_identity.py on structured expected and actual identities whenever the backend can export them. A script pass does not replace live readback; it proves only that the supplied identities agree.
8. Build a parametric skeleton and immediate preview
Model only overall dimensions, support footprint, primary and secondary masses, human contact zones, component envelopes, motion axes, interfaces, cables, thermal volumes, service space, and safety keep-outs. Immediately produce same-scale front, rear, side, top, bottom, and three-quarter views. Review proportion, silhouette, scale, posture, actions, interface count and direction, cable routing, center of mass versus support, assembly, and service.
Stop after the first massing preview. If the product still reads as stacked boxes, trays, exposed slabs, or unrelated primitives, return to architecture and form exploration. Do not continue by adding dimensions, labels, colors, or tertiary detail.
For consumer-product 3D, plan the complete review set before refinement: front, side, top, controlled three-quarter, key motion state, interface close-up, section, and exploded or hierarchy view. Mark a genuinely inapplicable view NOT_APPLICABLE with a reason; never omit it silently.
9. Co-develop structure, form, and process
Add one logical feature or subsystem at a time. Integrate wall strategy, stiffness, locating, fastening, sealing, clearances, motion, assembly sequence, tool access, service, routing, heat, and manufacturing direction. Add only enough detail to answer the current risk.
Keep appearance surfaces, structural surfaces, interfaces, and manufacturing datums traceable. Surface continuity, parting, draft, fasteners, seams, ports, and vents must support one design intent. Model ports and controls as real openings, recesses, flush panels, or justified proud features with purchased-part and use envelopes; a plate pasted onto a body is not interface integration.
10. Verify through separate evidence channels
Report separate dispositions for:
design intent and aesthetic coherence;
user flow, ergonomics, and access;
function, architecture, motion, and interfaces;
geometry, assembly, interference, and tolerance;
manufacturing, assembly, service, and cost risk;
safety and compliance architecture;
strength, stiffness, thermal, fluid, fatigue, or dynamic performance.
Use industrial-design-3d-quality.md, integrated-design-review.md, and the applicable engineering references. Without loads, materials, boundary conditions, rule thresholds, and verification methods, limit the conclusion to geometric plausibility or a labeled screening estimate.
11. Draft, deliver, and iterate
Create GB/T drawings only after the design reaches the maturity the drawing claims. Organize sources, exchange files, BOMs, evidence, and unresolved items with product-definition-release.md.
After feedback, update the brief, selection rationale, parameters, risks, aesthetic profile, and affected validation. Repair the smallest responsible source operation; never hide a source problem in a render or drawing.
When feedback reports poor appearance, poor practicality, misalignment, unequal repeated features, or another escaped defect, complete failure-learning.md. Record the escaped gate, root cause, containment, permanent process or Skill change, and next-version verification before claiming the issue closed.
At meaningful milestones or project closure, capture cross-window feedback with project-learning-loop.md. Keep raw personal records under ignored local-learning/; do not commit project paths, proprietary geometry, private images, or user-specific preferences to a public repository without explicit authority.
For repeated work on one product, preserve a project/branch/iteration lineage rather than overwriting prior attempts. Count repeated iterations as one project when generalizing preferences, but treat the same escaped defect across iterations as evidence that the owning gate or parameter source needs repair.
Automatic Local Learning
When ignored local-learning/settings.json exists with auto_capture: true, capture project learning automatically. Do not require the user to fill or copy a template when the current task already contains the required evidence.
Trigger on explicit like/dislike feedback, concept selection/rejection, accepted or rejected revisions, meaningful source redesign, gate escape, backend evidence failure, prototype/test disposition, milestone handoff, and project closure.
Build a schema-v2 record from the active project identity, exact user wording, design DNA, affected view/feature/state, expected versus observed result, source-level change, evidence, user disposition, unresolved scope, and privacy boundary.
Run scripts/record_project_feedback.py <temporary-record.json> --auto-lineage; let it assign the next branch iteration, parent record, and record ID.
Skip autosaves, formatting-only edits, unchanged repeated messages, and events without decision value.
Keep may_publish: false and never embed credentials or private asset content. Store controlled paths, IDs, hashes, or bounded descriptions.
Mention saved record IDs in the final summary. If automatic capture fails, disclose the failure and preserve the project work; do not pretend learning occurred.
Output Contract by Maturity
discovery package: brief, contradictions, TBDs, user flow, constraints, and opportunities.
concept package: at least three directions, common views, comparison, recommendation, and decision request.
design development package: selected direction, form grammar, parameter skeleton, key sections, interface, motion, assembly and service strategy, CMF intent, and risks.
industrial-design 3D package: revision-bound native model, parameter table, interface table, named motion configurations, same-camera review views, surface and rendering evidence, and user appearance disposition.
engineering candidate package: editable CAD, critical interfaces and dimensions, material and process basis, draft BOM, applicable analysis, and DRC or DFM results.
release candidate package: GB/T drawings, exchange files, configuration and BOM, verification evidence, TBDs, NOT_EVALUATED items, waivers, and external approval status.
In the final response, state the design decisions, their evidence, what the user confirmed, remaining assumptions, failed or unevaluated scope, current maturity, authoritative editable source, and the highest-value next test.
Deterministic Checks
Use these scripts when their input contract is available:
scripts/check_document_identity.py: compare requested and actual document, revision, configuration, units, and source identity.
scripts/check_pre_cad_brief.py: require the design-specific brief fields and immutable design DNA before detailed CAD.
scripts/check_dimension_authority.py: enforce A/B/C/D source grades and TBD behavior for critical interface geometry.
scripts/check_interface_alignment.py: compare 2D and 3D interface types, centers, sizes, orientations, and shared parameter source.
scripts/validate_motion_states.py: verify joint definitions, required named states, limits, collision, cable, and interface dispositions.
scripts/compare_render_viewset.py: verify source identity, camera invariance, image dimensions, nonblank output, and expected visual change.
scripts/check_model_manifest.py: bind required native, table, render, evidence, drawing, and exchange roles to one revision and source hash.
scripts/catalog_reference_library.py: create a stable read-only inventory for large local CAD and image libraries.
scripts/record_project_feedback.py: validate and append a privacy-bounded cross-window project-learning record to ignored local storage.
Treat every script as a narrow validator. A script pass never proves beauty, usability, physical performance, safety, manufacturability, or approval outside its declared fields.
Aesthetic Learning Contract
Treat user reference images as an evidence set, not as answers to copy.
Treat CAD block libraries as configuration, view, silhouette, and context references only. Their geometry, dimensions, brands, materials, dynamic actions, and annotations are not engineering authority until independently verified.
Extract repeated design logic across images and preserve counterexamples, context, and contradictions. One image does not create a permanent preference by itself.
Mark only explicit user likes and dislikes user-confirmed. Mark agent conclusions inferred with confidence.
Prefer transferring proportion, massing, line and surface logic, boundary strategy, detail rhythm, and CMF logic. Do not copy brand marks or distinctive trade identity unless the user explicitly requests it and has the right to use it.
After material aesthetic feedback, preserve the raw event with scripts/record_project_feedback.py. Store personal confirmed preferences in ignored local-learning/aesthetic-profile.local.md; update the public references/aesthetic-profile.md only when the user explicitly authorizes publication of a generalized, nonprivate rule. Do not claim persistent learning when no durable local record occurred.
