Willison Built a Datasette Map Plugin From a Single Claude Prompt
The datasette-maplibre 0.1a0 plugin draws Datasette table rows that contain coordinates onto a map. Simon Willison got nearly all of the code from one prompt to Claude Fable 5.2: study two existing plugins, merge them, and add ideas. The model itself proposed showing outlines as lightweight SVG and switching on the heavy WebGL map only when a row is visible. If you have working examples, you can hand an agent the task "combine and improve."
AI-processed from Simon Willison; edited by Hamidun News
Datasette is an open-source tool that publishes databases in the browser: you upload a table, and it gives you a ready-made interface with search and filters. If the table has coordinates, you naturally want to see them on a map, and that usually means choosing a library, wrestling with map tiles, and stitching different tools together. Simon Willison, the author of Datasette, handed almost all of this routine work to a neural network and published the result: the datasette-maplibre plugin, version 0.
1a0. It draws table rows on a map, both as points from latitude and longitude and as shapes from GeoJSON (a common text format for describing the borders of a country or a plot of land). Claude Fable 5.
2 wrote almost all of the code from a single prompt.
What the plugin does
Datasette already had two geospatial plugins. datasette-cluster-map plots points on a map from coordinates, and datasette-leaflet-geojson displays shapes from GeoJSON columns. The new plugin combines both ideas on a different foundation: it uses the MapLibre GL library and OpenFreeMap, a free vector tile map server supported by Zsolt Ero. Willison separately asks readers to sponsor Zsolt on GitHub.
In the demo on the page with capitals and countries, everything is visible on one world map. Countries are outlined and filled by continent, with colored points and cluster circles on top that show object counts. All 171 rows of the demo table fit on the map, and the legend counts them by continent: Africa 52, Asia 45, Europe 39, North America 16, South America 13, Oceania 6.
The panel lets you switch the mode to "Clusters," choose a column for coloring ("Colour by"), filter rows by the visible area, or download what is shown as GeoJSON. Clicking a point opens a window with the values from that row: for Caracas it shows the country, continent, the capital's population, and the coordinates. Below the map stays a regular table, where each country's border column has its own thumbnail.
How it was built with one prompt
For practitioners, the most interesting part is the method. The model was asked to study the code of the two existing plugins and Willison's local demo page with OpenFreeMap. It then had to build a new plugin that combines the capabilities of both. A request to be creative and come up with other useful features was added on top.
That last sentence paid off. Among the ideas, Willison highlights one: GeoJSON outlines in a table are first shown as lightweight SVG images, and the full WebGL view starts only when the row scrolls into view. This is a familiar engineering trick: on a page with hundreds of rows, you should not create hundreds of heavy graphics contexts at once. It was not specified in the brief; it appeared in response to an open-ended request.
Claude also built a detailed demo. It runs in the latest version of Datasette Lite, which is Datasette running entirely in the browser through Pyodide and WebAssembly. The link in the post loads the plugin and the demo database, and nothing needs to be installed. This makes it simpler for the plugin's author to show the work, and lets the reader check the result within a minute.
What to keep in mind
There are several caveats. Version 0.1a0 is an alpha, so the interface and behavior may change. Willison writes that he built "most of" the plugin with one prompt, so manual edits or additional requests are not ruled out, and he does not disclose the exact amount of follow-up work. The note contains no data on request cost, token usage, or how long the model took to run, so the economics of the approach cannot be assessed from it. Finally, this is a single example from an experienced author who knows the Datasette codebase well and knows how to set tasks for models.
The practical takeaway is still useful. When you have several working components and a demo page with the library you need, it pays to give an agent the task "study these examples, combine them, and add your own ideas." The model gets working context in the form of real code, not an abstract description. For developers of internal tools and data journalists, this is a way to quickly get a prototype, check it in the browser, and then take it to production by hand.
- Nov 21, 2026Anthropic Launches Linked Billing in Beta: Claude API Without Sharing the Key
- Nov 21, 2026Anthropic усилила «отчаяние» в Claude, и шантаж в тесте вырос с 22% до 72%
- Nov 21, 2026Anthropic Is Raising Up to $64.4 Billion in IPO as Orders Run Six Times the Shares on Offer
- Nov 21, 2026Banks Reject Anthropic IPO Orders from China and Hong Kong, Citing ITAR
Key Questions Answered
What the plugin does?
Datasette already had two geospatial plugins. datasette-cluster-map plots points on a map from coordinates, and datasette-leaflet-geojson displays shapes from GeoJSON columns. The new plugin combines both ideas on a different foundation: it uses the MapLibre GL library and OpenFreeMap, a free vector tile map server supported by Zsolt Ero. Willison separately asks readers to sponsor Zsolt on GitHub. In the demo on the page with capitals and countries, everything is visible on one world map.
How it was built with one prompt?
For practitioners, the most interesting part is the method. The model was asked to study the code of the two existing plugins and Willison's local demo page with OpenFreeMap. It then had to build a new plugin that combines the capabilities of both. A request to be creative and come up with other useful features was added on top. That last sentence paid off.
What to keep in mind?
There are several caveats. Version 0.1a0 is an alpha, so the interface and behavior may change. Willison writes that he built "most of" the plugin with one prompt, so manual edits or additional requests are not ruled out, and he does not disclose the exact amount of follow-up work. The note contains no data on request cost, token usage, or how long the model took to run, so the economics of the approach cannot be assessed from it.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.