AgentSkillsCN

dashboard-builder

与AI统计师协作，打造精美的数据仪表板。当收到主题简报、研究问题或一组DST表引用，并被要求制作仪表板时使用。推动完整流程：数据探索→仪表板制作→验证→视觉检查→迭代。通过提供描述主题、关键数据角度及期望仪表板概念的启动提示来调用此技能。

SKILL.md
--- frontmatter
name: dashboard-builder
description: >
  Collaborate with the AI statistician to create a polished data dashboard.
  Use when given a topic brief, research question, or set of DST table references
  and asked to produce a dashboard. Drives the full loop: data exploration →
  dashboard creation → validation → visual inspection → iteration. Invoke with
  a startup prompt describing the subject, key data angles, and desired
  dashboard concept.

Dashboard Builder

Mission

Drive an iterative conversation with the AI statistician (via the playground CLI) to produce a publication-quality dashboard on a given topic. Act as an editorial director: set the analytical direction, evaluate outputs, and push for improvements until the dashboard communicates its story clearly.

Execution flow

1. Prepare the brief

The user supplies a startup prompt. Extract from it:

  • •Subject — the topic or question the dashboard should answer
  • •Data angles — specific DST tables, indicators, or metrics to explore
  • •Dashboard concept — desired charts, comparisons, or narrative structure

If any of these are unclear, ask before starting.

2. Start a CLI session

bash
uv run python -m varro.playground.cli --user-id 1

Or resume an existing chat:

bash
uv run python -m varro.playground.cli --user-id 1 --chat-id <id>

Use interactive mode (not piped) so shell state persists across turns.

3. Explore the data (1–3 turns)

Send exploration prompts to the statistician. Guide it to:

  • •Discover which tables have the relevant data
  • •Run initial SQL queries to understand what's available
  • •Identify the right dimensions, time ranges, and breakdowns

Keep exploration focused. The goal is to gather enough understanding to plan a good dashboard — not to exhaustively analyze every angle.

4. Commission the dashboard (1 turn)

Send a clear dashboard request that includes:

  • •Dashboard name (slug)
  • •Which queries and outputs to create
  • •Layout structure (metrics row, tabs, grid)
  • •Filter dimensions (region, time period, etc.)
  • •Narrative framing — brief markdown text between charts

Be specific about the story the dashboard should tell. The statistician builds better dashboards when given a clear editorial direction.

5. Validate and inspect

After the statistician creates the dashboard:

  1. •Check the CLI output for any errors during creation.
  2. •Use :trajectory to inspect what the agent did if needed.
  3. •Read the created dashboard files directly to verify structure:
    • •queries/*.sql — check SQL correctness
    • •outputs.py — check output functions
    • •dashboard.md — check layout and narrative

6. Iterate (1–3 turns)

Based on inspection, send follow-up messages to fix or improve:

  • •SQL errors or empty results
  • •Missing or misleading chart labels
  • •Better filter defaults
  • •Additional tabs or metrics
  • •Narrative text that contextualizes the data

7. Final validation

Once satisfied:

  1. •Ask the statistician to run ValidateDashboard and Snapshot
  2. •Read the snapshot PNG to visually verify the dashboard
  3. •Read metrics.json to verify metric values make sense

8. Record results

After the dashboard is complete, note:

  • •The chat ID used (for trajectory reference)
  • •The dashboard slug and URL
  • •Any data caveats or limitations discovered
  • •Key editorial decisions made during iteration

Conversation strategy

Be directive, not vague. Instead of "make a dashboard about housing," say "Create a dashboard at /dashboard/boligmarked that shows property price indices by region since 2010, with a filter-select for region and filter-date for period. Include metrics for latest price index and year-over-year change. Use tabs to separate price trends from sales volume."

One major change per turn. Don't ask for five fixes at once. The statistician handles focused requests better than sprawling ones.

Read before asking for changes. Always read the current dashboard files before requesting modifications. Reference specific lines or output names.

Use the snapshot as ground truth. The PNG shows what users will actually see. If it looks wrong, that's what matters — not whether the code seems right.

Shell state reminder

Interactive CLI mode preserves dataframes across turns. The statistician can reference exploration results when building the dashboard. If you restart the CLI process, previous dataframes are lost — the agent will need to re-query.

Stopping criteria

Stop when:

  • •Dashboard validates without errors
  • •Snapshot PNG shows a clear, well-structured layout
  • •Metrics and charts tell a coherent story
  • •Narrative text contextualizes the data for a general audience