form

form

!checklist init ${CLAUDE_SKILL_DIR} --force

A form in the sculptor's sense is the three-dimensional presence of a thing — its weight, its surface, how the light falls on it. form is the depth-and-shape lens of the atelier suite: where color tokens say what hue lives here, form says how far above the surface it sits, what shape its corners make, and whether the surface has a texture or is empty. Its product is a written depth system: one light model with a tokenized elevation z-scale, a concentric radius family, and deliberately chosen gradients and texture — each committed as tokens before components consume them.

The governing fact: depth reads as physical when one light source drives four coordinated outputs; it reads as "CSS" when a lone fuzzy box-shadow is invented per component. The four outputs are the cast shadow (layered ambient + direct), the contact shadow (tight grounding layer), the top-edge highlight (inset light on the raised face), and the surface gradient (micro light-to-dark across the face). All four must share one light direction. Most UI depth work only touches the first — and the rest of the form system is the gap between "looks designed" and "looks premium." Radii follow the same logic: concentric nesting is not a visual preference, it is a geometric fact (the inner arc must subtract the padding to stay parallel); a nested card whose inner image has the same radius as the outer card does not look like the same system. And texture, like motion, earns its place only with a named job: depth, focus, or brand. Without a job, it is noise.

This is where the agent era bites:

• The agent invents a new box-shadow per component. Each is defensible alone; the set has four different light directions, three different blur radii, and two different coloring approaches. No individual value is obviously wrong; the system is incoherent. The light model must be a decision, written down, gated.
• The agent never does concentric radii. It assigns a border-radius by eye — or copies from a similar component — and the inner element gets the same radius as the outer container, so the corners visually pinch. This is the single most reliable fingerprint of unattended craft, and the one step that is almost never taken without a gate.
• The agent adds gradients because it can, not because they have a job. A slight gradient on a hero card, a subtle shimmer on a button, an ambient background glow on a dashboard — each seems harmless; together they are visual entropy with no signal. The question "what is this gradient's job?" must be a gate, not an afterthought.
• The agent never dithers. It writes a smooth gradient over a large area and ships banding. Blue-noise dithering is invisible when done correctly — and the banding is often invisible too, until a calibrated display or a side-by-side comparison makes it obvious. The check must be explicit.

Read references/one-light-model-and-elevation.md first — the technical foundation. The four outputs, the layered shadow stack, the z-scale, shadow coloring, dark-mode flip, performance rules. Load at the start; re-check at STAGE 0's gate.

Speak the user's language. The depth system is theirs to decide. Read their fluency and gloss terms on first use: elevation model / z-scale, ambient shadow / contact shadow / top-edge highlight, surface gradient (the face itself, not a decorative layer), concentric radius / inner = outer − padding, squircle / continuous curvature / G2-continuous, grain dithering / banding, interpolation space / oklch / oklab (for gradients). A system the user can't read is a style imposed, not shared.