Claude can write complete Datasette plugins now (simonwillison.net)

Claude Sonnet 4.5 (using Claude Code in “YOLO” mode) wrote an entire Datasette plugin end-to-end — code, tests, README and packaging — with minimal human prompting. The plugin, datasette-os-info, adds a /-/os JSON endpoint that collects and pretty-prints extensive OS details (platform, hostname, CPU/memory, Linux /etc/os-release fields, environment variables and the Python executable/version — in the author’s case Python 3.14). The author scaffolded the project with a datasette-plugin cookiecutter, ran tests (uv venv/uv run pytest), invoked Claude to implement and iterate on failures, built a wheel (python -m build), and published via GitHub Actions and Trusted Publishing to PyPI. A one-liner demonstrates live querying: uv run --isolated --with datasette-os-info datasette --get /-/os. This milestone matters because it shows modern coding agents can deliver fully functional, tested, packaged extensions with little human intervention, streamlining developer workflows (scaffolding → test → package → publish). It also highlights practical caveats: the model identified several sensitive fields (hostnames, env vars, paths) and suggested mitigations; the author added a README warning but kept the endpoint unauthenticated by default. Broader implications include faster plugin/app prototyping, stronger dependence on templates and training data examples, and new safety/supply-chain considerations when agents run with elevated permissions (the author used a permissions-skip mode).