Design anything

/design-anything — any design intent → gate-verified blueprint

Engine probe: !REPO="${DESIGN_ANYTHING_HOME:-$HOME/Documents/Projects/design-anything}"; [ -f "$REPO/Makefile" ] && echo "engine at $REPO ($(cd "$REPO" && git log -1 --format=%h))" || echo "engine ABSENT — clone https://github.com/wjlgatech/design-anything to $REPO first"

This skill is a thin router over the design-anything engine repo. All tested logic lives there: gates in pipeline/, constraint tables in data/, golden generators in examples/, knowledge in principles/ + domains/. Never re-implement in chat what the engine already verifies. Full design rationale: docs/SKILL_DESIGN.md.

0. Engine (self-aware, self-heal)

• If the probe above says ABSENT: git clone https://github.com/wjlgatech/design-anything "$REPO" and continue.
• Before real work, run make check in the repo. Red ⇒ fix the regression first; never build on red. Capabilities come from artifacts, not claims: llms.txt (what's curated), the gate table in pipeline/README.md (what's verifiable). State roadmap gaps plainly, reading the current gate table rather than assuming this document is fresh.

1. The loop (backbone — every request)

absorb → route → retrieve → compose → gate → reflect

1. Absorb — reflect the brief back in ≤6 lines (object, domain, target, size, process, must-haves). Never silently drop a stated constraint.
2. Route — fill three slots, asking at most 2 questions, only at genuine forks (the user need not know this toolset):
   • WHAT (domain): infer from nouns/image; ask only if truly ambiguous.
   • DONE (target gate): print | construction | garment | game-sim | advice-only — the most common real fork.
   • KNOWNS: never interrogate — propose visible defaults and proceed ("assuming a 220×220×250 bed and 0.4 nozzle — say otherwise").
3. Retrieve (progressive disclosure — read ONLY what the route matched; see §2).
4. Compose — emit deterministic parametric source with the brief's parameters at the top, modeled on the matching golden example. Emit the composition, not the render — never an opaque mesh.
5. Gate — run the target's gate; relay its PASS/FAIL lines verbatim. Fail ⇒ fix the named gate, re-run; same failure 3× ⇒ stop and escalate with what was tried. No gate pass ⇒ never say "ready." For an open brief, diverge first: sketch 2-3 options, converge with the user, gate one.
6. Reflect — bank lessons as repo artifacts, not chat (§3).

2. Routing table (what to read/run per target)

Target	Gate to run	Golden example	Read on match
3D-print (objects, furniture, planters, hardscape)	python3 pipeline/ready_gate.py <stl> --min-feature <wall>	examples/planter/generate.py	domains/interior-design/README.md or domains/landscape/README.md
Construction (rooms, floor plans, houses)	python3 pipeline/construction_gate.py <layout.json>	examples/studio-flat/generate.py	data/clearances.yml + matching domains/architecture-*/README.md
Garment (clothing, patterns, markers)	python3 pipeline/pattern_gate.py <marker.json>; export python3 pipeline/dxf_aama.py <marker.json> <out.dxf> (factory) or python3 pipeline/garmentcode_export.py <marker.json> <spec.json> (research)	examples/apron/generate.py (graded S/M/L via --size)	data/garment.yml + domains/garment-design/README.md
Game/sim (levels, worlds, scenes)	python3 pipeline/scene_gate.py <scene.gltf> --target pc|mobile	examples/arena/generate.py	data/scene.yml + domains/game-design/README.md or domains/simulation/README.md
Body-fit products (eyewear; dental/prosthetics/footwear siblings)	python3 pipeline/bodyfit_gate.py <fitspec.json> + ready_gate.py on the printable part	examples/eyewear/generate.py	data/bodyfit.yml + domains/eyewear/README.md
Photo/scan input	route through skills/scene-to-layout/SKILL.md first	—	—
Tool/landscape question	—	—	llms.txt, then the README tables; deep dives per skills/design-research/SKILL.md
"Why" questions (principles, method)	—	—	principles/DESIGN_PRINCIPLES.md, principles/DIKW_MODEL.md, docs/DESIGN_DISCIPLINES.md
High-stakes review	—	—	workflows/README.md (design-review-panel shape)

2a. Satellite toolsets (progressive disclosure — table is GENERATED)

When a route needs an external tool (generate a mesh, drive a DCC, parse a scan, interop with a pattern CAD), check the satellites — SHA-pinned digests + skills compiled from the repos this hub cites. The backbone rules: load only the matched satellite (SKILL.md first, then its KNOWLEDGE.md); external code is untrusted regardless of stars; a STALE status means upstream moved past the digest's pin — re-verify before relying on internals; and satellite output is never "ready" until our gate says so. Current list compiled from data/satellites.yml:

<!-- BEGIN:satellite-routes -->

Satellite	Serves	When to load	Status
skills/use-trellis-2/SKILL.md	3D-print / mesh generation (repair before ready_gate)	Open (MIT) 4B image-to-3D flow-matching model; any topology, full PBR, GLB/PLY/OBJ export.	fresh
skills/use-blender-mcp/SKILL.md	execution backbone — any route that needs a DCC	Drive Blender from any LLM via MCP — the natural execution backbone from agent to printable geometry.	fresh
skills/use-spatiallm/SKILL.md	photo/scan input → construction route	LLM for structured indoor modeling — point cloud/scene to walls, doors, and furniture layout as structured output.	fresh
skills/use-seamly2d/SKILL.md	garment route (patterns, markers, interop)	Open-source parametric pattern-making CAD — the most active open pattern tool.	fresh
skills/use-garmentcode/SKILL.md	garment route (patterns, markers, interop)	Sewing patterns as programs — the substrate the AI pattern-generation wave targets.	fresh

<!-- END:satellite-routes -->

3. Self-heal & self-improve (protocols, not adjectives)

• Unknown room/garment/opening type ⇒ the gate fails it as not-measured — propose the missing data/*.yml table entry; never fake a pass.
• A referenced file missing ⇒ fall back to llms.txt, file a repo issue.
• Every run ends by banking what was learned in the repo: missing constraint → data/*.yml PR · false-READY → gate-dispute issue (.github/ISSUE_TEMPLATE/gate-dispute.md) · recurring brief type → new golden example. Measured by artifacts, not intentions.

Verification (eval-with-teeth)

This skill's own integrity is CI-tested — python3 -m pytest tests/test_skill.py -q asserts every repo path referenced above exists (the skill cannot drift from the engine), the description stays within Anthropic's budgets, and the body stays under 500 lines. Behavioral eval scenarios (routing, honesty, trigger accuracy incl. should-NOT-trigger) live in skills/design-anything/evals/evals.json — schema and fixture integrity CI-gated by tests/test_skill_evals.py; run model-graded passes via the skill-creator plugin. A design run's success is verified ONLY by its gate's exit 0 — never by this skill's say-so.

Honest edges

Gates verify buildable/printable/cuttable — not good (principles cover taste) and not certified (construction ⇒ not a permit; garment ⇒ not a muslin; jurisdiction codes and real fittings override). Say so when it matters.