Design studio
Multi-agent frontend design workflow for new surfaces, full redesigns and high-value visual iteration. Separates product framing, visual direction, implementation and blind browser evaluation; roots the target before work starts; preserves immutable iterations; resumes from recorded evidence; optionally consumes Impeccable detector output; and documents the accepted visual system. Use Review for audit or polish without redesign.From its SKILL.md
npx -y skills add George-RD/design-studio --skill design-studioAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
10.3 KB, as published. Nobody here has run it
Design Studio
Design Studio protects visual judgement from implementation anchoring. It is a controlled workflow, not a style prompt collection.
Non-negotiable boundaries
| Role | May see source | May see prior scores | Owns decisions |
|---|---|---|---|
| Planner | yes | no | scope and success criteria only |
| Visual Director | no | no | visual proposals only |
| Builder | yes | no | implementation only |
| Evaluator | no | no | observations and scores only |
| Orchestrator | as needed | yes | SELECT / REFINE / PIVOT / SHIP / HALT |
- Visual Director never receives HTML, CSS, JSX, component names, selectors, implementation diffs or an unattended assignment index.
- Evaluator never receives source, implementation effort, the full design description or prior scores. It judges the live render against product and surface success criteria.
- Builder may add semantics, accessibility, responsive behaviour and required states. It may not replace the selected direction with a safer one.
- Orchestrator is the sole decision owner. Evaluator output cannot contain a workflow decision.
Lanes
| Lane | Use when | Authority |
|---|---|---|
| Studio | New surface or material redesign | workflow.yaml |
| Review | Audit or polish without a new visual world | references/review/polish.md |
| Design system | Codify the selected build or extend a proven system | Codify step in workflow.yaml |
| Meta | Improve the workflow itself | references/meta.md plus run traces |
Routing
- Create, build or redesign a page or product surface: Studio.
- Overhaul an existing path or URL: Studio, with baseline capture.
- Audit, polish, slop, accessibility, hierarchy or ship readiness without redesign: Review.
- A single component or narrow CSS correction: ordinary implementation or Review, not a full Studio run.
- Extract tokens and design DNA from an accepted build: Design system.
When a request mixes audit and redesign, Studio owns the task unless the user explicitly asks for a report first.
Context and run integrity
Load references/context.md and references/runtime-integrity.md before planning.
roots.jsonrecords repository, app and context roots with evidence.capabilities.jsonrecords the browser, runnable target, detector, question and copy tools that actually exist.events.jsonlis the append-only step journal used for resume.PRODUCT.mdis durable product truth.COPY.mdis optional durable voice, claim and terminology guidance.DESIGN.mdis the proven visual system. It is authority for extensions; during an explicit redesign it is evidence and an anti-reference unless the user says to preserve it.surface-brief.mdholds only the current surface job, action, content, proof and constraints.selected-direction.mdis the source-free visual contract summary for one iteration.
Inspect before asking. Ask only for material gaps. Mark unattended inferences as assumptions. Never invent customers, prices, benchmarks, capabilities or testimonials.
Studio execution
Run workflow.yaml end to end.
- Root and probe: resolve repository, app and context roots; probe required and optional tools before spending an iteration.
- Plan: confirm product truth, classify the surface as
persuade,operate,readorexperience, capture a baseline for overhauls and choose a finite build budget. - Assign: for unattended work, commit a reproducible seed and candidate slot before directions are generated. Keep both hidden from Visual Director.
- Explore and select: Visual Director produces three equally specified, materially different directions. A pinned direction, the user or the precommitted slot selects one.
- Direct and build: Visual Director writes the source-free contract. Builder implements it in a new immutable iteration and records fidelity evidence.
- Mechanical preflight: use
references/quality-gates.md. Each scan is a complete current snapshot. Mechanical checks can block craft or functionality but never assign visual quality. - Blind evaluation: Evaluator interacts with the live page at verified desktop and mobile viewports, captures zones and records observations and scores without a workflow decision.
- Decide: Orchestrator applies the ordered decision table. REFINE keeps the world. PIVOT replaces it. SHIP moves to tiered final selection.
- Finish and accept: eligible, mechanically clean iterations outrank higher averages with a failed criterion. A fresh reviewer checks the selected result. Apply at most one correction batch, then write an acceptance receipt for the final tree.
- Codify: copy the accepted build to
harness-output/site/; writeDESIGN.md, design DNA, tokens and the report from final evidence.
Budget selection
| Class | Builds | Use |
|---|---|---|
| quick | 2 | focused surface or straightforward page |
| standard | 4 | ambitious page or product screen |
| ambitious | 6 | complex experience, many states or high visual risk |
An explicit budget is clamped to 1–8. A PIVOT consumes a build. Budget exhaustion selects the best available result and labels it honestly.
When Studio starts without a runnable browser target, it records the limitation, reduces the budget to one build, runs mechanical preflight and halts without selecting a visual winner.
Resume
A run resumes only by explicit run ID or an already active run root.
- Validate completed event receipts and their artifacts.
- Continue from the first incomplete or invalidated step.
- Never rerun a completed build in the same iteration.
- Never infer completion from a directory that merely exists.
- Preserve failed and superseded evidence.
Artifacts
Every run is immutable below harness-output/runs/<run-id>/.
run.json
roots.json
capabilities.json
events.jsonl
spec.md
sprint-contract.md
surface-brief.md
scores.json
baseline/
iterations/<n>/
direction/direction-assignment.json
direction/directions.md
direction/direction-selection.json
direction/selected-direction.md
direction/design-description.md
site/
serve.json
design-flags.json
mechanical-findings.json
screenshots/
observation.json
critique.md
finish/
selection.json
selected-site/
selected-serve.json
selected-direction.md
corrected-site/ # only when correction runs
corrected-serve.json # only when correction runs
correction-verdict.json # only when correction runs
final-tree.json
acceptance.json
Only after acceptance may Orchestrator replace harness-output/site/. Git history is not the iteration store.
Review lane
Execute references/review/polish.md without workflow.yaml.
- Resolve roots and probe the runnable target first.
- Deterministic checks own source and browser-computed facts such as exact contrast, overflow and token drift.
- Visual lenses own visible hierarchy, composition, rhythm and generated-template feel.
- With no browser, or when either required viewport cannot be reached, return
visual_status: unverifiedand the mechanical report. - With a browser, run one inspection batch, apply one grouped fix batch when requested and confirm once.
Quality semantics
- The brief wins. A pinned aesthetic or system is not penalised merely because a detector recognises the pattern. Use a documented waiver.
- Common is not automatically bad. A familiar component can be correct for an operational task; it simply does not create originality by itself.
- Redesign replaces; refinement preserves. Preserve product truth, copy claims, behaviour and explicit brand commitments. Do not average old and new visual worlds together.
- Mechanics are a floor, not a direction. Zero detector findings do not make a page distinctive.
- Findings are current. A rerun replaces the current open set. History may explain a resolution, but it cannot keep a fixed problem open.
- Finish from evidence. Record the design system after the final build proves it.
Required references
| Need | File |
|---|---|
| Product, copy, design and surface context | references/context.md |
| Roots, capabilities, resume, assignment and acceptance | references/runtime-integrity.md |
| Machine workflow, paths and decisions | workflow.yaml |
| Visual Director prompt | agents/design-agent.md |
| Builder constraints | references/generation.md |
| Mechanical gate and Impeccable integration | references/quality-gates.md |
| Customer-facing copy changes | references/copy.md |
| Blind visual evaluation | agents/evaluator.md |
| Review-only audit | references/review/polish.md |
| Overhaul baseline | references/overhaul.md |
| Workflow tuning | references/meta.md |
Prerequisites and degradation
Studio needs file I/O, shell access and isolated subagents. A complete visual decision also needs a runnable target and browser automation. Verify requested viewport widths with window.innerWidth or the adapter equivalent.
- No browser or runnable target in Studio: preserve one build and current mechanical snapshot, then halt unselected.
- No browser or missing required viewport in Review: return mechanical evidence with visual status
unverified. - No Impeccable: use the fallback gate and record
detector: fallback. - No user answer mechanism: use the precommitted deterministic assignment.
- No image generation: present equal text directions; do not fabricate comps.
- No
business-copy-style: applyreferences/copy.mdand recordcopyWorkflow: local-rules.
Extending the workflow
- Add a capability as one focused leaf plus one routing row. Do not create another always-loaded design skill.
- Keep style examples out of always-loaded prompts unless the brief names them.
- Keep scoring language about qualities, not favoured aesthetics.
- Version
SKILL.md,workflow.yaml, plugin metadata and eval metadata together. - After a material model change, test whether each layer still improves outcomes before retaining its cost.