Design DNA

Most "make it look like this" requests fail the same way: the reference is looked at once, an impression is formed, and the impression is coded from memory. What survives is a vague resemblance — the palette drifts, the rhythm flattens, and the one effect that gave the reference its character is missing entirely.

This skill inserts an artifact between looking and building. The reference is first deconstructed into a Design DNA JSON, and only that JSON is used to generate. The intermediate step is what makes the result checkable: every colour, radius, and easing curve in the output traces to a recorded field, so a mismatch is a diff rather than an argument about taste.

Provenance: the three-dimension taxonomy and several extraction rules are adapted from the MIT-licensed zanwei/design-dna skill. See .claude/rules/moai/NOTICE.md for the retained copyright notice.

The three dimensions

The split matters because the three are extracted differently and fail differently.

Collapsing them loses information in both directions. A token dump with no style dimension reproduces the colours and none of the character; a mood board with no system dimension reproduces the vibe and no two elements align.

Field-by-field schema and enum vocabularies: references/dna-schema.md.

Phase 1 — Structure

When the request is for the schema itself ("what does a design profile contain?"), read references/dna-schema.md, present the three dimensions and their field groups, and ask whether any dimension should be extended or dropped before extraction begins.

Phase 2 — Analyze (reference → DNA JSON)

Read references/dna-schema.md first, then work reference by reference.

• Image or screenshot → read it directly and analyse its visual properties.
• URL → fetch the page. Prefer the source: <canvas> elements, WebGL contexts, animation-library imports, custom shaders, IntersectionObserver scroll triggers, and SVG <animate> are stated facts, where a screenshot only supports inference.
• Video or interaction capture → the only reliable source for motion timings, scroll choreography, and transition personality.

Extraction rules

1. Colour roles are assigned by dominance, not by hue. Primary is the colour holding the most area, secondary the supporting surface, accent whatever carries the calls to action. Map the neutral scale as an ordered ramp from the lightest background to the darkest text.
2. Measure radius relatively. Record it against the element height that carries it — an absolute 12px is meaningless once the button is a different size, while "half the control height" survives rescaling. The concentric-radius rule for nested surfaces is moai-ref-ui-polish's; do not restate it here, apply it there.
3. Infer type scale from ratios, not from absolute sizes. Heading-to-body ratio and line-height rhythm transfer across viewports; pixel sizes do not.
4. Density is proximity. Judge spacing from element gaps relative to the base unit, and section rhythm from whether those gaps stay constant or escalate between sections.
5. When references conflict, record the dominant pattern and name the variant. Averaging two references produces a design neither of them has.
6. Fill every field. An empty string is indistinguishable from "not looked at". Where the reference genuinely does not exercise a field, say so in the value — an explicit "not observed" is data; a blank is a gap.
7. Effects absent from the reference get enabled: false. This is the non-invention rule, and it is load-bearing: an unset flag invites the generator to add a particle field nobody asked for.
8. What cannot be identified goes in composite_notes. A screenshot shows that a surface glows without showing how. Describing the observation beats guessing the implementation and recording the guess as fact.

Output the completed JSON, then ask whether any value should be adjusted before generation.

Phase 3 — Generate (DNA JSON + content → artifact)

Read references/effects-implementation.md before implementing any visual_effects entry.

Order matters, because early decisions constrain later ones. Colour and typography together carry most of a design's identity, so they are settled first; effects are layered onto a design that already works without them.

1. Colour and typography
2. Spacing and layout
3. Shape and elevation
4. design_style qualitative fields — these guide the judgement calls the token values do not determine
5. visual_effects
6. Motion and interaction, last: an interface whose static layout is wrong is not rescued by animating it

Emit design_system as CSS custom properties in a single :root block, so every downstream value has one definition and a token swap is one edit.

Fetch real assets rather than approximating them. When the reference is a URL and the design needs its logo, image, or font, retrieve the actual asset from that source. A recreated approximation is the single most visible difference between a copy and a clone.

Default output is a self-contained file — inline CSS and JS, no build step — unless a framework was specified. The self-contained convention itself is documented once in moai-domain-html-report § output; the difference here is scope: that skill renders a report from markdown, this one generates a designed artifact from a DNA profile.

Profile persistence

Phase 2 can end by saving the completed DNA JSON as a named profile, and Phase 3 can start from the active profile instead of re-extracting the same reference: references/diagram-profiles.md documents the mechanism — the project-root .design-dna/ store, the marker-first active selector, the slug grammar, schema validation with "not observed" backfill on load, and the confirm-before-overwrite / verify-by-re-read save path. A project with no profile marker proceeds exactly as above; persistence adds a save hook and a start-from option, never a routing change.

Delivery gate

Verify before handing over. Each item below has failed in practice.

• Every colour in the output traces to a DNA palette entry.
• Font families, spacing rhythm, and radius match their DNA tokens.
• Body text meets 4.5:1 contrast, large text 3:1.
• prefers-reduced-motion: reduce is honoured — not merely detected.
• Nothing renders for an effect whose enabled flag is false.
• Animation loops use requestAnimationFrame; setInterval is not an animation primitive and drifts against the display refresh.
• Canvas and WebGL contexts are sized to their container and handle resize.
• The declared fallback_strategy is actually implemented, not just recorded.

Polish iteration

A first pass that reads thin is usually not a token problem — it is an attention problem, and re-reading the reference beats re-reading the output. Re-attach the same references and audit against them along six axes: hierarchy, ornamentation, typographic rhythm, motion, materiality, and overall interface finish. Merge the conclusions back into the current implementation rather than regenerating from scratch, which discards the parts that were already right.

Boundaries

This skill owns one thing the others do not: turning an existing reference into a structured profile, and generating from that profile. Where the design is authored from first principles rather than deconstructed, those skills own it.

References

• references/dna-schema.md — the three-dimension field list and enum vocabularies
• references/diagram-profiles.md — named-profile persistence: the .design-dna/ store, marker-first resolution, slug grammar, load-time validation
• references/effects-implementation.md — performance tiers, technology selection, and per-effect implementation patterns