Chart and Dashboard Planner

A dashboard is a set of answers wearing a layout, and most bad ones fail before the first pixel: nobody wrote down the questions. This skill plans charts and dashboards backwards from decisions — what the viewer needs to decide, which comparison answers it, which visual form carries that comparison — and produces a build-ready spec for whatever BI tool or charting library renders it.

When to use this skill

• Designing a new dashboard for a team, an executive audience, or an operational review
• A dashboard exists but nobody uses it — or everyone exports the data to look at it elsewhere
• Choosing the right chart form for one specific finding in a report or deck
• Reviewing a dashboard proposal before build time gets spent on it

Workflow

1. Collect the questions, not the metrics. Interview the audience or infer from their decisions: "do we need to intervene on churn this week?" is a question; "churn rate" is merely a column. Write down three to seven questions, ranked. A dashboard serving more than about seven questions is several dashboards sharing a URL.
2. Classify each question by its comparison, because form follows comparison:
   • Change over time → line for trends, bars for discrete periods, area only for parts-of-whole over time
   • Ranking across categories → horizontal bars, sorted by value, never alphabetically
   • Part-to-whole at one moment → stacked bar; a pie only for two or three slices; beyond that, bars
   • Distribution → histogram or box plot; when in doubt, show the shape
   • Relationship between two measures → scatter, with a trend line only if you mean to claim one
   • Single value against expectation → a big number with target, delta, and a thin strip of history
3. Decide the state each chart must make legible at a glance: good, bad, or watch. That means targets, thresholds, or prior-period ghosts drawn on the chart itself — a line with no reference is a Rorschach test.
4. Lay out by scan order. Top-left carries the sharpest "are we okay?" signal — headline numbers with deltas; trends sit beneath; diagnostic breakdowns live below the fold. Group by question, never by data source. One screen per audience-moment: scrolling is for diagnosis, not for status.
5. Specify interactivity honestly: the two or three filters the audience will genuinely use (time range, segment), a drill path from each headline to its diagnostic, and the default state — the dashboard must answer its top question with zero clicks.
6. Design the annotation layer: release markers, definition changes, seasonal notes. A dashboard that cannot explain its own anomalies generates meetings instead of preventing them.
7. Define empty, loading, and broken states. A stale dashboard that looks fresh is worse than a broken one; every view carries its data-as-of timestamp somewhere the eye lands.
8. Dry-run before build. Sketch with real numbers from last month and put the top question to a target viewer. If they cannot answer within ten seconds, rework the plan before anyone writes a query.

Output format

## Dashboard spec: <name> — audience: <who>, cadence: <when they look>

| # | Question (ranked) | Comparison | Chart form | Reference shown | Drills to |
|---|-------------------|------------|-----------|-----------------|-----------|

Layout: <sections in scan order, one line each>
Filters: <the few that earn their place>     Default state: <what loads>
Annotations: <event sources>     Data-as-of: <where displayed>
Out of scope: <questions this dashboard refuses to answer>

Guardrails

• One question per chart; a chart answering two questions answers neither
• Bar axes start at zero, always; line axes may zoom but must say so on the axis
• No dual y-axes for unrelated units — use small multiples instead
• Color encodes meaning (state, series identity), never decoration; one accent color means "look here"
• Every number on the dashboard traces to a defined metric spec; no orphan calculations invented in the chart layer
• If the honest answer is a table, ship a table — sorted, with the key column emphasized

Ten-second test checklist

• Top question answerable with zero clicks, inside ten seconds
• Every chart shows a reference: target, prior period, or threshold
• Sort orders are by value or by time, never accidental
• A stranger can tell when the data was last updated
• Nothing on screen exists merely because the data was available