Visant Refine

Finishing. This skill assumes the decision to include the element was already made and defended elsewhere, and only answers how it is drawn.

The hierarchy, and it is not negotiable

Refine never justifies an element. If the answer to "which variable does this move" is silence, the fix is to delete, not polish. A beautiful chart of data nobody uses is slop with better finishing, and it costs more attention than the ugly version.

Work order, always in this sequence:

1. Pass the visant-frontend cut test. Name the variable per element.
2. Inventory what already exists. See "Before drawing" below.
3. Pick the form. Load dataviz before the first line of chart code.
4. Draw with the shared anatomy (table below).
5. Only then decide the animation, with references/smart-motion.md.
6. Verify the way visant-frontend says (390px with real data, backend down, keyboard), and run both gates:

node ~/.claude/skills/visant-refine/scripts/refine-scan.mjs <target>   # finishing layer
node ~/.claude/skills/visant-killer/scripts/killer-scan.mjs --report AUDIT.md <target>

--report is not optional: killer-scan refuses to print GATE OPEN without the surface table and the ## Screen seen section (## Tela vista also works; generate the skeleton with --init-report AUDIT.md). And a zero from refine-scan alone is not a complete pass: a clean detector feels like finishing because it ends in an exit code, and that is exactly how visant-frontend gets skipped entirely.

refine-scan has six rules, all for things that compile, pass lint and still break: raw color in a chart prop (only in files that import recharts), chart without a state wrapper, unscoped defs id, easing or duration outside the SSoT, transition-all and ease-in in UI, and width or height animation. The state wrapper counts if the file mentions ChartState, isEmpty, noData, semDado or vazio, so it works in any repo. Findings in components/ui are suppressed (upstream primitive) and counted at the end, never hidden. --self-test proves the six rules both ways before you trust a zero.

Before drawing: discover what the repo already has

No new primitive without searching first. Almost every refine request is composition, not construction, but you only know after looking. Run at the root of the target repo, in this order:

# 1. installed chart lib (recharts, visx, nivo, chart.js, echarts, d3)
grep -E '"(recharts|@visx/[a-z-]+|@nivo/[a-z-]+|chart\.js|react-chartjs-2|echarts|d3)"' package.json
# 2. existing chart state wrapper and tooltip
grep -rlE 'ChartState|ChartContainer|ChartTooltip|isEmpty' --include=*.tsx --exclude-dir=node_modules .
# 3. slider, ratio bar and skeleton primitives
find . -path ./node_modules -prune -o -type f \( -iname 'slider.tsx' -o -iname 'progress.tsx' -o -iname 'skeleton.tsx' \) -print
# 4. motion tokens (easing and duration SSoT)
grep -rnE -- '--(ease|dur|duration)-[a-z0-9-]+:' --include=*.css --exclude-dir=node_modules . | head
find . -path ./node_modules -prune -o -type f -iname 'motion.ts*' -print
# 5. color token format (bare HSL channels or full color)
grep -nE -- '--(primary|muted-foreground|destructive):' $(find . -path ./node_modules -prune -o -name 'globals.css' -print)

Write down what you found in a "need | exists at" table before the first line. For whatever is missing, search in this order:

1. The repo itself (the commands above).
2. The Visant library, ui.visantlabs.com, already-refined pieces from any project: gauge, sim-slider, measure-bar, activity-heatmap, streamgraph, count-up, elastic-slider, skeleton-rail. Install with npx shadcn@latest add @visant/<name>, with "@visant": "https://ui.visantlabs.com/r/{name}.json" in components.json; each one's spec sheet at /r/<name>.md.
3. The official shadcn primitive (npx shadcn@latest add slider progress skeleton).

Design system components are never created or edited without explicit permission. Warn before installing. Primitive default overrides go in the consumer.

The shared anatomy

What makes a screen look designed is not each element being beautiful, it is all of them telling the same thing in the same order. Every finishing element carries these four roles, always in the same visual sequence:

Rules for all four, in any element:

• Color is the exception. One neutral, and color only where there is a real exception. Text uses a text token, never the series color. Color follows the entity, never the ranking position.
• No number without a ruler. A value without a qualifier is trivia. "R$ 25.201,42" says nothing; "11,3% da receita bruta" does.
• No denominator, no ratio. Zero denominator is null and the row does not render. "0,0%" states a false fact.
• A shape is only valid inside its domain. A ratio bar has scale while the value fits the limit; past it, the bar saturates and lies, and the number saying how far it went replaces the bar. This is not a detail, it is the rule that shows up most.
• Empty, error and loading are three states, and all three are designed.
• Same family at both ends: when adding the second element of a kind, extract the first. Two variants of the same thing is the start of three.

Element catalog, each with anatomy, props and its typical mistake: references/elements.md.

Charts

Load dataviz and follow its procedure. What changes here, and is not there:

• Series color comes from a token, and the token format is checked before the first line (command 5 above). Where the CSS stores bare HSL channels (--primary: 0 0% 9%, the older shadcn default) the prop needs the wrapper: stroke="hsl(var(--primary))". Bare var(--primary) yields the string 0 0% 9%, which is not a color, and the stroke disappears with no error in tsc, lint or detector. Raw Tailwind palette and hex in JSX are out too: a class like text-purple-400 slips past generic detectors.
• A state wrapper surrounds every chart (the repo's ChartState, or the equivalent command 2 found). An empty framed axis is dataviz's "0 items".
• clipPath and defs ids collide by name. Two charts on the same page with the same id overwrite each other; scope per instance.
• Limits and goals are a ReferenceLine, labeled, not a background color. The line is the fact; the color would be opinion.
• description and heavy payloads stay out of the chart data.

Ready recharts recipes (limit with ReferenceLine, cumulative crossing a band, axis-less sparkline, ratio gauge) and the token-by-function map: references/charts.md.

Slider simulation

A slider that projects a scenario is a lying machine if the design does not separate projection from fact. The minimum contract:

1. The real value stays on screen, always, marked on the track itself, and the label says which is which. Two numbers of the same unit and different kinds without labels corrupt the reading.
2. Stops at the limits that matter. The slider stops at the numbers that change the decision (a regime ceiling, a tax bracket), not at round multiples.
3. The band is drawn on the track, so the user sees where the regime changes before dragging.
4. Back to real is one click, and it only shows while a simulation is active.
5. Nothing persists. Simulation does not write to the database, and the URL only carries state if sharing is part of the request.
6. Keyboard moves it, aria-valuetext speaks the formatted value.

Full contract, with the state pattern and the classic mistakes: references/simulators.md.

Control panels

An inspector or editor sidebar is the same anatomy at row scale: label left, value right in tabular-nums, the unit as qualifier one tone dimmer, the slider fill as evidence. What makes one look engineered instead of assembled:

• The control takes the shape of the data. 2–5 options are a segmented control, a point in 2D is a 3 × 3 grid, a vector is one row of fields, a gradient is its bar with stops on it, a curve is the curve plus its numbers. Cover the labels: if the shapes still say what they edit, it is designed.
• The chrome takes the shape of the state. Reset appears in a folder header only after a value changed; a dependent control exists only while its switch is On; rare actions sit behind one ⋯; help is on demand, not a caption.
• Depth by line, not by box. Sticky folder headers, a dashed divider for nesting, quiet rows and loud values.
• The mode comes before what it changes, and every number takes drag, typing and arrow keys.

A dev-only tweak panel is an install (Leva, Tweakpane, DialKit), not a build. Full rules, the data-to-control table and the DialKit-vs-fork case that measured them: references/control-panels.md.

Smart animation

Smart means chosen by the variable, not by opportunity. The rule that generates the others: animate what carries the decision, at the moment the decision changes.

• A value the user is producing by hand (dragging the slider) does not animate: it follows the finger. Animating here is latency that looks nice.
• A value that changes as a consequence (the total the slider recalculated, after release) may count up to the new number, short, once.
• Crossing a limit is an event, and an event deserves weight: the band's state change is the animation, not a glow on the card.
• Nothing animates on mount if the user already read that number in the previous navigation.
• A hundred times a day does not animate.

Decision table per case, with SSoT values: references/smart-motion.md.

Worked case: tax panel

A tax simulation panel is the canonical example of all three topics together, written step by step (what went in, what was refused and why) in references/tax-panel-case.md. Read it before refining any number panel, because the mistakes are the same: a saturating bar, projection confused with fact, and animation on the wrong data.