Design orchestrator

Design Orchestrator

The single entry skill for design work. It locates the task on the intent × deliverable map and delegates to one of 9 specialist spokes. The cross-cutting model every design task shares — research and constraints before generation, the decision ledger, and the anti-AI-slop quality gate — lives in design-core; read it before generating any pixels.

Routing map (intent → spoke)

Design a real product UI (default path)

• Any interface, landing page, dashboard, redesign, or visual polish → refero-design (research is mandatory — reference before implementation)
• Need styles / palettes / font pairings / chart types per stack → ui-ux-pro-max (searchable design-intelligence DB)
• Want a specific, opinionated look applied → swiss-design (Swiss/modernist system in Tailwind)

Engineer the interface to a senior bar

• Metric-based rules, component architecture, motion, hardware-accelerated performance → taste-skill
• Emit a reusable, agent-enforceable spec file (anti-generic standards) → stitch-design-taste (produces DESIGN.md)

Run a design-agent workflow

• Iterate drafts / branch variations / extend to more pages from a component → superdesign

Theme an existing artifact

• Apply or generate a theme for slides, docs, reports, or HTML → theme-factory

Produce static visual art

• Poster, art piece, print-ready .png / .pdf → canvas-design

Pull brand identity from the web

• Extract logos, colors, backdrops, brand name from a URL → openbrand

Standard Operating Flow

1. Locate the task on intent × deliverable: is the output a product UI, a spec, a themed artifact, static art, or brand data?
2. If the output is a product UI or anything user-facing, pull the gate from design-core first — research → lock references → ledger decisions → generate → check against the slop gate. Skipping research is the failure mode.
3. Delegate to the spoke(s). Multi-step asks fan out in deliverable order (e.g. "brand this landing page" → openbrand for tokens → refero-design for the layout → taste-skill for the build).
4. Return: chosen spoke(s), the references/constraints that anchored the work, the deliverable type, and the next action.

Guardrails

See design-core. In short: research before generation — never average from the model's defaults; ground every choice in a concrete reference and record it in the decision ledger. Keep type scale, spacing, and color on a system, not ad-hoc. Treat "make it look good" as a request to find references and commit to a direction, not to improvise. The cluster's value is craft that resists generic AI output — don't quietly fall back to the default look.

Loading spokes on demand

To keep CLI startup context lean, this cluster's spokes are not separately registered as skills — only this orchestrator and its *-core are enumerated. When you route to a spoke named above, load it on demand by reading its file:

~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md (or skills/<spoke-name>/SKILL.md inside the skill-clusters repo).