Openchart

Data Visualization with OpenChart

Note: SVG design capabilities (logos, icons, graphics) have moved to the opendesign plugin. Install with /plugin install opendesign@opendata-skills.

Source of truth: load the types first

Before authoring or modifying a spec, load the OpenChart type definitions. They are the canonical source for field shapes, enums, and defaults. This skill complements the types — it does not replace them. When this skill and the types disagree, the types win.

Try these locations in order; stop at the first one that resolves:

1. Installed package (most common): node_modules/@opendata-ai/openchart-core/dist/index.d.ts. The full chart, table, graph, sankey, tilemap, and map spec surface is rolled into this single bundled .d.ts.
2. Source repo (if you’re working in the openchart monorepo): packages/core/src/types/spec.ts and packages/core/src/types/layout.ts. JSDoc comments here are richer than the bundled .d.ts.
3. Published CDN (no local install, network available): https://unpkg.com/@opendata-ai/openchart-core/dist/index.d.ts (redirects to the latest published version).
4. GitHub raw (if unpkg is unreachable, or you want the richer source JSDoc): https://raw.githubusercontent.com/tryopendata/openchart/main/packages/core/src/types/spec.ts. For version-exact types, pin to a release tag instead of main: https://raw.githubusercontent.com/tryopendata/openchart/core-v<version>/packages/core/src/types/spec.ts, matching <version> to the installed @opendata-ai/* package version. Note main may document unpublished surface.
5. Fallback: if none of the above are reachable (no filesystem, no network fetch), use the type sketches in this skill and flag the uncertainty in your response so the user knows you authored without canonical types.

The names worth grepping for once you have a types file open: ChartSpec, TableSpec, GraphSpec, SankeySpec, TileMapSpec, GeoMapSpec (+ GeoMapGeo, GeoMapPointsLayer), MarkType (the 16-mark union), Encoding, EncodingChannel, MarkDef, Chrome, Metric, EndpointLabelsConfig, Annotation (union), TextAnnotation, RangeAnnotation, RefLineAnnotation, LegendConfig, LabelSpec, SeriesStyle, AnimationSpec, ThemeConfig, A11yConfig, SeriesSearchConfig, YouDrawItConfig.

Rendering via MCP

If you have access to the OpenData MCP openchart tool, use it to render specs interactively – it handles rendering, theming, and responsive layout. Two ways to call it:

• Quick chart: pass chartType (bar, line, area, point, arc, lollipop, circle, rect, tick) plus the field names (x, y, color). Good for the common x/y shapes.
• Full control: pass a spec object (the OpenChart JSON described in this skill) – required for marks the quick chartType list doesn’t cover (range, waffle, calendar, parliament, beeswarm, text, rule) and for any encoding/chrome/annotation options.

Feed data with sql (the tool re-runs your previous query_sql – preferred, don’t copy rows) or inline data rows (keep under 1000). The tool’s own description carries version-pinned links to the full schema for constrained generation.

When the openchart tool is not available, output the spec as JSON for the user to render with <Chart> / createChart() (see rendering reference (opens in a new tab)).

Structured outputs (tool-use / constrained generation). OpenChart ships a published JSON Schema and a generated llms.txt, both derived from the spec types so they never drift:

• JSON Schema for constraining an LLM tool call to a valid spec. Import from the core ./schema subpath (not the type barrel): @opendata-ai/openchart-core/schema is the full VizSpec union (usable directly as an Anthropic tool input_schema); @opendata-ai/openchart-core/schema/chart.schema.json is the chart-only subset covering all 16 marks; .../table.schema.json is the table subset. Files also live at packages/core/schema/*.schema.json and on unpkg/GitHub raw for fetch-only agents.
• llms.txt at the repo root (https://raw.githubusercontent.com/tryopendata/openchart/main/llms.txt) is the compact narrative surface: install, core concept, the mark-encoding table, and validation notes. Prefer it as a quick primer when you can’t load the full .d.ts.

Validate before rendering. validateSpec(spec) from @opendata-ai/openchart-engine returns { valid, errors, normalized }. Each error carries a machine-readable code, the offending path, and a repair-friendly suggestion. Field references are checked against the columns in the provided data, and a misspelled field name gets a Levenshtein-based Did you mean "..."? clause pointing at the nearest real column – use that clause to auto-repair typos in a generate-validate loop.

What this skill carries that the types don’t

The types tell you the shape of a valid spec. This skill carries the things types can’t express:

• Behavioral defaults that live in the normalizer/compiler, not in the type signature. (E.g. bar and area both default to stacked when colored; grouped/overlap behavior requires explicit stack: null.)
• Cross-field interactions and truth tables. (E.g. how legend.show, endpointLabels, and the legacy end-of-line labels interact.)
• Runtime semantics. (E.g. legend toggle recompiles, rebalances the y-axis, locks the color scale, hides per-series UI, refuses to hide the last visible series.)
• Editorial judgment. (E.g. when to pick a bar vs a line, color strategy, annotation density, data-resolution budgets.)
• Worked examples for common stories.
• Failure modes and gotchas that you’d only learn after rendering and looking at the result.

If you find yourself restating a field shape this skill already documents, prefer the types — they’re authoritative and won’t drift.

Core concept: Write a VizSpec JSON object, render with <Chart> / <DataTable> / <Graph> / <Sankey> / <TileMap> / <GeoMap> (React/Vue/Svelte) or createChart() / createTable() / createGraph() / createSankey() / createTileMap() / createGeoMap() (vanilla JS). The engine validates, compiles, and renders. Specs are plain JSON, no imperative drawing. See https://github.com/tryopendata/openchart (opens in a new tab) for the rendering engine.

CSS is required. OpenChart’s stylesheet must be loaded for proper rendering (chrome, tables, tooltips, brand watermark). Framework imports handle this automatically, but CDN/standalone HTML needs an explicit <link>:

<link rel="stylesheet" href="https://esm.sh/@opendata-ai/openchart-vanilla/styles.css">

See rendering reference (opens in a new tab) for details.

Chart Selection Decision Tree

Single KPI / KPI row         -> spec.metrics: Metric[] (label+value+delta, rendered between subtitle and chart)
Temporal x-axis column?      -> 1 series: line | 2-5 series: line + color | 6+: filter to top 5
Categorical + numeric?       -> Ranked list: bar (horizontal) | Periodic (Q1, Jan): bar (vertical) | 2-6 composition: arc
Two numeric columns?         -> point (optional size/color for 3rd/4th dims)
Categorical + series + num?  -> stacked bar (use color for series)
Distribution/spread?         -> circle (strip plot) | many observations per group: beeswarm
Change between two values?   -> range (dumbbell / arrow / floating bar per category)
Part-to-whole as counts?     -> waffle ("x of 100" unit grid)
Election / legislature seats?-> parliament (hemicycle) | half-donut result: arc + startAngle/endAngle
Daily value over a year?     -> calendar (GitHub-style heatmap)
Nodes + edges / network?     -> graph (force/radial/hierarchical layout)

Visualization Types

Each type has a detailed reference with full spec, encoding rules, and examples. Load the reference when you need the details.

Bar orientation: The engine infers orientation from encoding. x: nominal/ordinal + y: quantitative = vertical (column-style). x: quantitative + y: nominal/ordinal = horizontal bar. Override with mark: { type: "bar", orient: "horizontal" | "vertical" }.

Arc variants: mark: "arc" renders a pie chart by default. Add innerRadius > 0 to get a donut: mark: { type: "arc", innerRadius: 40 }. For an election-style half-donut, restrict the sweep with startAngle/endAngle in radians (mark: { type: "arc", innerRadius: 40, startAngle: -Math.PI/2, endAngle: Math.PI/2 }); the engine resizes a partial sweep to fill the chart area so a half-donut isn’t drawn at half size.

New-mark behavior the types don’t tell you:

• range needs the second value channel: x + x2 (horizontal, the common editorial form with y as the category) or y + y2 (vertical). style picks the form: "dumbbell" (default, muted start dot + accent end dot + connector), "arrow" (arrowhead at the x2/y2 end, strongest “change over time” read), or "bar" (plain floating range bar). colorByDirection: true colors increases with the theme’s positive color and decreases with negative; a field-based encoding.color wins over it. Use this mark for dumbbell/change plots rather than faking one with two overlaid point series.
• waffle takes color (the category) and a quantitative share via theta (the same part-to-whole channel arc uses; y is still accepted as a deprecated alias). units (default 100) sets total cells, columns (default 10) the grid width. Shares normalize to units via largest-remainder rounding so cells always sum exactly; a small nonzero share can round to 0 cells (there’s no minimum-one-cell floor), but it still appears in the legend.
• calendar takes x (temporal, one row per day) and color (quantitative per-day value). Multi-year data stacks one band per year sharing a single color scale. Date math is UTC, so "2024-01-15" parses as UTC midnight. Days with no data render as empty achromatic cells, distinct from the scale minimum. weekStart (“monday” default / “sunday”) sets the top row; cellRadius rounds the cells.
• parliament takes color (party) and seat count via theta (y is still accepted as a deprecated alias). Only shape: "hemicycle" ships (concentric semicircular arcs). Parties fill left-to-right in data order, so sort your rows by political spectrum yourself. majorityLine (default true) draws the threshold line and “N to win” label; seatRadius defaults to "auto".
• beeswarm takes one quantitative positional channel (the value axis) plus an optional nominal channel for grouped lanes; size scales dot area. The cross axis is pure pixel-space with no scale, so tall stacks can overflow a short container – cap the size range or give it vertical room.

Collapsed mark types: These mark aliases no longer exist as separate values. Use the canonical marks instead:

Reference Routing

Always load when generating a new chart spec: encoding-channels.md (opens in a new tab), format-strings.md (opens in a new tab), color-strategy.md (opens in a new tab)

Common reference bundles:

• New chart: encoding-channels + format-strings + color-strategy + editorial-writing + (mark-specific ref)
• Design polish: design-review + visual-qa + editorial-writing
• D3 infographic: d3-core-patterns + infographic-design + (topic-specific D3 ref)

Spec discriminant

• Charts use mark (16 marks: "bar", "line", "area", "point", "circle", "arc", "text", "rule", "tick", "rect", "lollipop", "beeswarm", "range", "waffle", "calendar", "parliament").
• Tables, graphs, sankey, tilemap, geo maps use type ("table" | "graph" | "sankey" | "tilemap" | "map").

For the full top-level shape — every optional field, exact enum values, defaults — load ChartSpec, TableSpec, GraphSpec, SankeySpec, TileMapSpec, GeoMapSpec from index.d.ts (see “Source of truth” above).

Behavior worth knowing that the types don’t tell you:

• animation is off by default. Set true for sensible per-mark entrance defaults; pass an AnimationConfig for per-phase control. See animation.md (opens in a new tab).
• crosshair is off by default and only renders on line/area charts.
• endpointLabels is auto-on for ≥2-series line/area and auto-suppresses the traditional legend in that case. See the “Endpoint Labels” section below for the full suppression truth table.
• hiddenSeries on the spec hides series on first render. The vanilla adapter also maintains a separate runtime hidden set populated by legend clicks; that triggers full engine recompile (y-axis rebalance, locked color scale, per-series UI hide). See “Legend Toggle (Runtime)” below.
• display: 'sparkline' strips chrome, axes, legend, watermark, animation, and crosshair for inline KPI-card use. Explicit per-field overrides still win (set chrome.title and you’ll still get a title in sparkline mode).
• seriesSearch (boolean | { placeholder }) renders a typeahead “find your country” input over a categorical color encoding; selecting values highlights them (multi-select chips). Mutually exclusive with edit mode (search wins).
• youDrawIt ({ from, prompt?, revealLabel?, comparisonLine? }) is the NYT “draw your guess before the reveal” format. Line marks only, single-series only; mutually exclusive with edit mode and seriesSearch. The vanilla instance exposes resetDrawing() / revealDrawing() and an onReveal(guess) callback.
• mark.fillPattern: 'auto' layers a per-series SVG pattern (hatch, dots, crosshatch) over each fill so filled marks (bar/area/arc) stay distinguishable without color vision; 'none' (default) is solid fills.
• Accessibility is automatic: openchart generates alt text, ARIA labels, and a hidden screen-reader data table. Override the alt text with top-level description (Vega-Lite sugar) or ( wins); set to a purely decorative chart. The engine also emits console warnings when adjacent series or text fall below WCAG contrast, naming the nearest passing color.

Chrome elements (chrome.eyebrow / title / subtitle / source / byline / footer / brand): each takes string | ChromeText. The eyebrow is a tracked, accent-tinted kicker above the title. The brand is a right-anchored block on the footer row paired with a small accent dot — setting it suppresses the default tryOpenData.ai watermark. All chrome text supports \n for explicit line breaks and auto-wraps at the container width. Exact field shape: see Chrome and ChromeText in index.d.ts.

KPI metric row (charts only): metrics: Metric[] at the top level (not inside chrome) renders a horizontal row of label+value cells between the subtitle and the chart area. Each cell can carry a delta and a secondary value. Auto-stripped in sparkline mode and at narrow/short containers, or when value text would overflow its cell. Exact field shape: see Metric in index.d.ts.

Mark (Charts Only)

mark is either a string (one of the 16 marks: "bar", "line", "area", "point", "circle", "arc", "text", "rule", "tick", "rect", "lollipop", "beeswarm", "range", "waffle", "calendar", "parliament") or an object — see MarkDef in index.d.ts for the full field set (type, point, interpolate, orient, innerRadius, startAngle, endAngle, fill, stroke, strokeWidth, opacity, fillPattern, plus per-mark fields like style/colorByDirection for range, units/columns for waffle, weekStart for calendar, shape/seatRadius/majorityLine for parliament, etc.).

Behavior the types don’t tell you:

• Bar orientation is inferred from encoding types. x: nominal/ordinal/temporal + y: quantitative = vertical column. x: quantitative + y: nominal/ordinal = horizontal bar. Override with mark: { type: "bar", orient: "horizontal" | "vertical" } only when the inference is wrong.
• mark: "arc" is a pie by default. Pass innerRadius > 0 to get a donut.
• mark.fill accepts a GradientDef for linear or radial gradients. Gradients can also appear as conditional color values in encoding.color. See gradients.md (opens in a new tab).
• Default entrance animation depends on mark type when animation: true. Bars clip-reveal from the baseline; lines draw progressively; areas draw + fade; arcs scale from center; points pop in with scale + fade; text/rules/ticks fade in. See animation.md (opens in a new tab) for per-mark defaults and per-phase overrides.

Collapsed mark aliases (don’t use these):

Examples:

{ "mark": { "type": "line", "point": true, "interpolate": "monotone" } }
{ "mark": { "type": "arc", "innerRadius": 40 } }
{ "mark": { "type": "bar", "fill": { "gradient": 

Encoding Channels (Charts Only)

Charts map data to visuals via encoding channels: x, y, color, size, detail, key, x2, y2, opacity, strokeDash, angle, text, tooltip, theta, facet. Each channel is an EncodingChannel (field, type, aggregate, axis, scale, bin, timeUnit, sort, format, title, stack, condition, value). For the full shape and enum values: load Encoding and EncodingChannel from index.d.ts.

Behavior the types don’t tell you:

• field + type is the minimum for any channel. type must be one of "quantitative" | "temporal" | "nominal" | "ordinal" — picking the wrong one is the most common spec authoring bug (see Spec Anti-Patterns below).
• axis.format is d3-format with a literal-suffix extension (e.g. ".1f%"). The literal suffix is OpenChart’s add-on; native d3 doesn’t support it. See format-strings.md (opens in a new tab) and the Format Strings section below for the percent-form pitfall.
• scale.nice: true is the default for quantitative and temporal scales — it rounds the domain outward to clean tick values. On temporal scales this can shift the domain by years; set nice: false when you need precise control.
• stack defaults to stacked ("zero") when colored for both bar and area; line = n/a. Set stack: null for grouped/overlap behavior. See the per-mark default table in encoding-channels.md (opens in a new tab).
• See encoding-channels.md for conditional encoding examples and the full per-channel guide.

Data Transforms (Charts Only)

Apply transforms (filter, bin, calculate, timeUnit, aggregate, fold, window) to data before encoding. Transforms run in array order. Filters support data-relative time references for temporal fields. See data transforms reference (opens in a new tab).

Layer Composition (Charts Only)

Overlay multiple marks in a single chart using layer. Each layer is a standalone spec with its own mark, data, and encoding.

{
  "layer": [
    { "mark": "bar", "data": [...], "encoding": { ... } },
    { "mark": "line", "data": [...], "encoding": { ... } }
  ]
}

Layers share the same coordinate space. Use this for combo charts (bar + line), adding reference lines, or overlaying annotations.

Dual-Axis Charts (Independent Y-Scales)

When two series have incompatible value ranges (e.g., revenue in millions vs. headcount in thousands), use resolve: { scale: { y: "independent" } } on the layer spec. Layer 0 gets the left y-axis; layer 1 gets the right y-axis. Both share the x-axis.

{
  "resolve": { "scale": { "y": "independent" } },
  "layer": [
    {
      "mark": { "type": "bar", "opacity": 0.85 },
      "data": [

Dual-axis rules:

• Max 2 layers – there are only left and right y-axes.
• Both layers must have compatible x-field types (both ordinal, both temporal, etc.).
• Use axis.labelColor on each layer’s y-encoding to color the axis labels to match the series – this is the standard dual-axis pattern (Datawrapper, Highcharts style).
• Use labels: { density: "none" } on both layers to avoid label collisions between the two series.
• The engine zero-aligns both y-scales so zero sits at the same pixel height on both axes. Annotations target the primary (left) y-scale.
• Works with any combination: bar + line, bar + area, area + line.

Legend Configuration (Charts Only)

legend accepts position, show, columns, symbolLimit, maxRows, offset, exclude. Full shape: LegendConfig in index.d.ts.

Behavior: Position is responsive by default — the engine picks top / right / bottom / bottom-right / inline based on container width. Set position to override. Use show: false when the legend is redundant (e.g. bar charts where the y-axis already labels each category). For multi-series line/area, leaving show unset triggers auto-suppression in favor of the endpoint chip column — see the truth table below.

Endpoint Labels (Multi-Series Line/Area)

For multi-series line/area charts, the engine renders a column of chip+swatch labels at the trailing edge of each series (rounded pill with a colored bar swatch + label + last value). Auto-on for ≥2 series; off for single-series and non-line/area.

endpointLabels accepts boolean | EndpointLabelsConfig (show, valueField, format, width, showMarker, showLeader, markerStyle). Full shape: EndpointLabelsConfig in index.d.ts.

Suppression truth table (≥2-series line/area). The traditional legend, the endpoint column, and the legacy end-of-line labels are three knobs that interact. The implementation lives in packages/engine/src/legend/suppression.ts and is the single source of truth — if this table ever drifts, the engine wins.

Single-series charts: column hidden by default (nothing to identify).

Common patterns:

• Default multi-series: leave both unset – you get the endpoint column, no traditional legend.
• “I want a top legend instead”: legend: { position: 'top' }, endpointLabels: false.
• “Both legend and endpoint column”: legend: { show: true } (endpointLabels stays auto-on).

Legend Toggle (Runtime)

Clicking a legend entry hides/shows the corresponding series at runtime. This goes through engine recompile (not CSS hide), so:

• The y-axis rebalances to the remaining visible series.
• The color scale stays locked – remaining lines keep their original palette colors (engine injects a stable scale.domain from the unfiltered data).
• Per-series UI hides with the line: endpoint chip, leader, dot annotation, and any text annotation anchored to that series.
• The last visible series can’t be hidden (the toggle is a no-op).
• Range annotations and reference lines pass through unchanged (they anchor to constant axis values, not series).

Pass onLegendToggle to observe these clicks; you don’t need to wire up hiddenSeries yourself for default behavior. Use hiddenSeries on the spec to start with specific series hidden on first render.

Label Density (Charts Only)

labels accepts density, format, prefix (full shape: LabelSpec in index.d.ts). The choice that actually drives the chart is density:

For multi-series line/area, prefer endpointLabels (the chip+swatch column) over density: "endpoints" (the legacy fallback). The chip column wraps long names and resolves collisions; the legacy labels reserve a large right margin for long series names.

Format Strings

Both axis.format and labels.format accept d3-format (opens in a new tab) strings plus a literal suffix extension. See format strings reference (opens in a new tab) for the full table.

Critical: When data is already in percentage form (12.5 meaning 12.5%), use ".1f%" not ".1%". The d3 % type multiplies by 100, so 12.5 becomes 1,250.0%.

Per-Series Styling (Charts Only)

Use seriesStyles: Record<string, SeriesStyle> to override visuals for individual series, keyed by the color-field value. Use this for “highlight one series, dim the rest” patterns or for a dashed reference series alongside primary data. Field shape: SeriesStyle in index.d.ts. Editorial guidance + examples: series-styles.md (opens in a new tab).

Data Resolution

Keep data arrays under 150 rows per series. More data doesn’t make a better chart - it makes a slower, harder-to-read one. Reduce resolution before building the spec, not after.

How to reduce: When querying APIs, aggregate before passing data to the spec:

• Use group_by=year with aggregate=avg(value) to go from monthly to annual
• Sample every Nth row for evenly-spaced data
• Filter to the time range that matters (don’t chart 75 years when the story is about the last 10)

Multi-series charts are multiplicative. A 3-series line chart with 300 points per series = 900 data rows. Reduce each series to ~50-80 points for a clean result. For the same 25-year span, annual data (25 points x 3 series = 75 rows) reads better than monthly (300 x 3 = 900 rows).

Why this matters beyond readability: Large data arrays inflate the spec JSON, slow rendering, and generate massive accessibility tables in the DOM. A 900-row chart produces a ~19,000px tall screen-reader table that can break page layout if styles don’t load correctly.

First Draft Checklist

Run these checks before outputting a spec. These catch the issues that most often require iteration after rendering.

Spec Anti-Patterns

For design anti-patterns (titles, color, annotations), see design review (opens in a new tab).

Known Gotchas

Rendering and component behaviors that aren’t obvious from the spec alone.

Custom D3.js Infographics

When a visualization goes beyond what declarative specs can handle (creative metaphors, unusual layouts, treemaps, generative art), fall back to raw D3.js + SVG. Note: sankey, tilemap, and geo maps (choropleth + symbol) are first-class types with their own spec formats (see references/sankey.md (opens in a new tab), references/tilemap.md (opens in a new tab), and references/map.md (opens in a new tab)), and scrollytelling is now first-class too – use the chart story API (base spec + patch steps) instead of hand-rolling D3 scroll effects (see references/story.md (opens in a new tab)). Use the D3 reference only for heavily customized layouts. These references cover D3 implementation patterns: