See the workflow · Install · Roadmap · Issues
Design Studio is a portable design-engineering Agent Skill for creating, reviewing and judging visual work while keeping source code out of visual direction and evaluation.
The Builder sees the implementation. The agents choosing and judging the design do not. Code can implement the idea; it does not get to choose it.
| Lane | Use it for | Primary authority |
|---|---|---|
| Studio | New interactive surfaces and material redesigns | skills/design-studio/workflow.yaml |
| Review | Audit and polish while preserving the existing visual world | skills/design-studio/references/review/polish.md |
| Document | Quotes, invoices, SOWs, proposals, reports and other paginated print/PDF artifacts | skills/design-studio/references/document/document.md |
Studio and Review also have optional Claude Code slash-command adapters. Document is invoked through the Agent Skill directly; there is no separate Document slash command.
| Room | Roles | What it can see |
|---|---|---|
| Source room | Planner, Builder | Product files, source, routes, assets and implementation constraints |
| Blind room | Visual Director, Evaluator | Product truth, screenshots, rendered pages, the live page and scoped critique |
| Decision owner | Orchestrator | Run evidence, budgets, findings and preserved iterations |
The Evaluator reports visible evidence. It cannot decide to refine, pivot or ship. The Orchestrator applies the recorded rules.
With the Agent Skill installed, ask a capable host to use Design Studio directly:
Use Design Studio to create a landing page for a small coffee shop.
Use Design Studio to review ./site without redesigning it.
Use Design Studio to design a paginated proposal and judge the rendered pages.
Claude Code can expose Studio and Review through its optional adapter:
/design-studio:create a landing page for a small coffee shop
/design-studio:create --overhaul ./site --goals "keep the information architecture" replace the visual direction
/design-studio:review ./site
The canonical, host-portable product is the Agent Skill:
npx skills add George-RD/design-studionpx skills is the installer used to obtain the Agent Skill; it is not a Design Studio runtime dependency. The public GitHub repository is the install source, so no separate skills.sh registration is required. A capable host can also copy skills/design-studio/ into its local skills directory.
The Claude Code plugin is an optional, thin adapter over the same Agent Skill and does not reduce supported capability when absent on a capable host. See ADR 0003 and optional adapters.
/plugin marketplace add George-RD/design-studio
/plugin install design-studio@design-studio
/reload-plugins
Claude CLI equivalents: claude plugin marketplace add George-RD/design-studio and claude plugin install design-studio@design-studio.
- File and shell access plus isolated subagents.
- Studio needs a runnable target or valid
serve.jsonand browser automation for a complete visual decision. - Review needs rendered interactive evidence for a verified visual verdict.
- Document needs complete ordered rendered-page evidence from an existing artifact or a host-supplied renderer for a visual decision.
When required visual evidence is missing, Design Studio preserves bounded evidence and reports the result as unselected or unverified rather than inventing a verdict.
A standard Agent Skills install contains skills/design-studio/. Root commands/, agents/ and .claude-plugin/ are optional compatibility adapters.
Repository scripts/, benchmarks/, test/ and .github/ are research and development support, not runtime dependencies. Deterministic helpers needed by a supported run ship inside the skill behind its internal runtime contract. After installation, the copied skill runs from that contract and host capabilities without requiring the skills CLI. See the runtime boundary.
roots.jsonanchors the repository, app and product context before work starts.capabilities.jsonrecords the tools and evidence that were actually available.events.jsonlmakes each step resumable without overwriting completed work.- Directions, builds, rendered evidence, findings and decisions stay under one immutable run folder.
- Mechanical checks and source-blind visual judgement remain separate.
- Only an accepted final tree becomes durable visual authority.
COPY.md is optional. Growth Arsenal may provide separate approved offer or copy inputs through the composition contract; it is not required to install or run Design Studio.
harness-output/
├── runs/<run-id>/
│ ├── roots.json, capabilities.json, events.jsonl
│ ├── spec.md, sprint-contract.md, surface-brief.md
│ ├── iterations/<n>/
│ └── finish/
├── site/
├── design-system/
└── report.md
Design Studio v1.7 has one self-contained Design Studio runtime: the installed Agent Skill. It ships a local deterministic mechanical runtime under skills/design-studio/runtime/; supported behavior does not branch on whether another design package is installed.
External systems are research inputs, not runtime dependencies. Impeccable and Emil Kowalski's skills are credited sources, not install requirements. Adopted methods live in the local kernel with provenance and progressive disclosure. Growth Arsenal remains a separate optional skill and composes only through the neutral artifact contract. See ADR 0002.
Codex and Claude Code are exercised through the standard Agent Skills install in CI. The first-class Document lane is renderer-neutral and uses complete rendered pages as visual evidence. See the prepared v1.7 evidence record.
Mechanical findings remain evidence. They never substitute for source-blind judgement of the rendered result.
MIT. Attribution and licence notes for Impeccable and Emil Kowalski's skills are recorded in NOTICE.md; neither is required to install Design Studio.