Data Viz

AI-First Data Visualization

Philosophy

Build production-quality interactive data interfaces with modern component libraries — no vendor lock-in, embeddable anywhere. When no tool is specified, build code-first. When the user explicitly names a BI tool, use it — only suggest code-first if they ask for options or hit a technical blocker.

Technology Stack

Full API patterns & code: references/component-guide.md

Framework Priority

1. React + Tailwind — Default when JSX/TSX supported
2. HTML + CSS + Vanilla JS — Fallback (use D3 or Chart.js)
3. Python (Plotly/Dash) — Python-only environments only

Library Selection

Supporting: Tailwind CSS · Radix UI · Framer Motion · Lucide React · date-fns · Papaparse · lodash

Building a Visualization

Step 1: Understand the Data Story

Before code, identify: What question does the data answer? Who is the audience (exec → KPIs only, analyst → drill-down, public → narrative)? What's the ONE key insight? Design around it.

Step 2: Choose Chart Type

Step 3: Build the Interface

Start from this layout — remove what the data doesn't need:

┌─────────────────────────────────────────┐
│ Header: Title + Description + Date Range│
├─────────────────────────────────────────┤
│ KPI Row: 3-5 metric cards + sparklines  │
├─────────────────────────────────────────┤
│ Primary Visualization (largest chart)   │
├──────────────────┬──────────────────────┤
│ Secondary Chart  │ Supporting Chart/Tbl │
├──────────────────┴──────────────────────┤
│ Detail Table (sortable, filterable)     │
└─────────────────────────────────────────┘

A single insight might just be one chart with a headline and annotation. Scale complexity to audience.

Step 4: Design Principles

• Data-ink ratio: Remove chartjunk — unnecessary gridlines, redundant labels, decorative borders
• Color with purpose: Encode meaning (red=bad, green=good, blue=neutral). Max 5-7 colors. Single-hue gradient for sequential data
• Typography hierarchy: Title → subtitle (muted) → axis labels (small) → data labels
• Responsive: min-h-[VALUE] on all charts. Grid stacks on mobile
• Animation: Entry transitions only, duration-300 to duration-500. Never continuous
• Accessibility: aria-label on charts, WCAG AA contrast, don't rely on color alone
• Dynamic color safety: When colors come from external sources (brand palettes, category maps, API data, user config), never apply them directly as text color without a contrast check. Dark colors are invisible on dark card backgrounds. Safe pattern: use the external color only for non-text elements (left border, dot, underline); always use the standard text color (white / var(--text)) for the label itself. If color-coded text is required, apply a minimum lightness floor: color: hsl(from brandColor h s max(l, 60%))
• Icon semantics: Verify every icon matches its label's actual meaning, not just its visual shape. Common traps: using a rising-trend icon (📈) for metrics where lower is better (latency, error rate, cost); using achievement icons (🏆) for plain counts. When in doubt, use a neutral descriptive icon over a thematic one that could mislead

Step 5: Interactivity & Annotations

Priority: Tooltips (every chart) → Filtering → Sorting → Drill-down → Cross-filtering → Export → Annotations

Annotations turn charts into stories. Mark: inflection points, threshold crossings (amber), external events (indigo/red), anomalies (red), achievements (green). Limit 3 per chart. Implementation: references/component-guide.md → Annotation Patterns.

Step 6: Tell the Story

• Headline states insight: "Revenue grew 23% QoQ, driven by enterprise" — not "Q3 Revenue Chart"
• Annotate key moments directly on chart
• Contextual comparisons: vs. prior period, vs. target, vs. benchmark
• Progressive disclosure: Overview first — detail on demand

Environment-Specific Guidance

Anti-Patterns

• Screenshot/static charts — build interactive components
• Defaulting to BI tools unprompted — build code-first when no tool specified
• Default matplotlib — always customize in Python
• Rainbow palettes — use deliberate, meaningful colors
• 3D charts — almost never appropriate
• Pie charts > 5 slices — use horizontal bar
• Unlabeled dual y-axes — use two separate charts
• Truncated bar axes — always start at zero
• Filtering or mapping over a field not confirmed to exist in the data export — an undefined field in .filter() or .map() produces empty arrays or NaN silently, not an error; always validate the exported schema matches what the chart code consumes

Pre-Delivery Checklist

Before marking a dashboard complete:

• Every tab / view activated — all charts render (no blank canvases, no unexpected 0–1 axes)
• Every field referenced in chart/filter code confirmed present in the data export
• All text readable on its background — check explicitly when colors come from external data
• All icons match their label's meaning
• Tooltips appear on hover for every chart
• No chart silently receives an empty dataset — add a visible empty state or console warning
• Mobile: grid stacks correctly, no body-level horizontal overflow