GitHub 39
Skip to content

Data explorer

Your agent explores a dataset and turns what it finds into an interactive page of findings, graphs, matrices and tables, from one report.json.

Open the example: an agent built it with this skill from the public Nuxters data, in one run.

The agent starts from the data, not the page. It writes down the questions you want answered, profiles every field, and computes every number in a prep script, so a rerun on new data rebuilds the page. Then it writes report.json: the data rows, the filters, and the blocks on each tab. A shared template draws it, so the agent writes no HTML, CSS or page script. Every page gets the same parts:

  • a top bar with the tabs and a light and dark theme
  • a strip of numbers that stays on every tab, and the rows behind them at the bottom
  • flat sections of charts, segmented bar lists, heatmap matrices and tables, two to a row
  • a filter sidebar with search, the period, and one list per entity, where every name keeps one color across the page and a click filters every block

The build rejects text past its limits and names the fix, so the page stays numbers first. It then screenshots every tab in light and dark, and the first on a phone, and the agent reads each one before it hands the page back.

The result is one HTML file. It opens from disk or any static host, and you can publish it to Drop to share it privately, with a new version on each rebuild and comments on the numbers.

Ask with the data in hand, such as "What does this export say about who contributes to our repos?" or "Make this CSV explorable by country and month." To be sure, say "Use the data-explorer skill."

  • You explore data and filter it yourself: this skill.
  • You choose between options or approve a plan: Report.
  • Findings about something that already exists: Review document.

Node with playwright installed, for the build that checks the page against its limits and screenshots it. With comark-ui installed too, the page carries Vega and opens offline. Without Node, as in a chat app, the agent fills the skill's page template itself, and the page loads whatever it couldn't fill, such as Vega, from jsDelivr when it opens.

  • The first tab opens with two or three numbers that answer what you asked, each with one sentence.
  • Every tab mixes block types: a table, a matrix or a bar list next to its charts.
  • A name has the same color in the sidebar, the tables and the charts, and clicking it filters the page.
  • Definitions and caveats sit on the Method tab, not under every chart.
  • The agent hands back the screenshots it read and what the data couldn't answer.