Willison Builds a Map Plugin With Claude
Executive Summary
- Simon Willison says he built most of a new Datasette mapping plugin with one prompt to Claude Fable 5.2. The prompt pointed to two existing plugins and a local demo, giving the model working examples of the features and integration it needed.
- Datasette 1.0a43 adds streaming exports and security fixes as Willison uses Claude to work through its release backlog. One fix addresses exposure of private stored queries’ titles and SQL; maintainers using that feature should review the release’s advisory.
A Map Plugin Built From Working Examples
Willison’s November 15 datasette-maplibre 0.1a0 release combines MapLibre GL with OpenFreeMap’s vector-map service. It displays geographic data from Datasette, including GeoJSON shapes and latitude/longitude points on the same map.
His prompt told Claude to inspect datasette-cluster-map, datasette-leaflet-geojson and a local OpenFreeMap demo, then build a plugin combining their features. He also asked it to “be creative and think outside the box.” One result he liked: lightweight SVG previews of geographic shapes, with WebGL previews loaded only when they scroll into view.
Claude also built a demonstration that runs inside Datasette Lite, using Pyodide and WebAssembly. Readers can try the plugin and sample database entirely in a browser. That makes the result easier to inspect than a screenshot or an account of how quickly code appeared.
The reusable technique is straightforward: give a coding model reference implementations for the application’s conventions, the desired features and the unfamiliar library. Willison’s account supplies a working example of that approach, though no controlled comparison of development time.
The accompanying Datasette 1.0a43 release shows the maintenance work alongside the new plugin. Its Response.stream() method lets output plugins stream rows; built-in CSV and TSV exports now use it. The release also fixes a bug that could reveal a private stored query’s title and SQL, without allowing the query to execute. Willison reports that the project’s 1.0 milestone is 53% complete.
Further Reading
- Jack Clark’s Import AI 481 — Connects agents’ ability to spend money on ML research with the difficulty of overseeing their work, linking the underlying research.
- OpenAI’s candidate lattice definition of the Standard Model — OpenAI credits GPT-6 Astra with helping develop a detailed proof; an ambitious scientific claim worth following for independent assessment.