Simon Willison wrote the 4.0rc2 release candidate for sqlite-utils—a Python library for working with SQLite databases—using almost exclusively Claude Fable. 37 prompts, 34 commits, 1,321 lines added and 190 removed across 30 files. Reported cost: around $149. He shared the complete process, transcript, and pull request.

The point isn’t that the agent writes code—we already knew that. The point is it found critical bugs Willison himself had missed, like a delete_where() method that never committed transactions and poisoned the connection (data loss). Fable categorized 5 blocking issues to fix before release, wrote docs for the new transactional model, and sorted out compatibility with Python 3.12’s autocommit flag. Willison then had everything reviewed by GPT-4o, which caught two more problems—proof that cross-checking different models works.

It’s not magic: Willison steered every step, reviewed every commit, and spent hours on the final PR. But the grunt work—analysis, writing, testing, documentation—the agent did it, while he went to the July 4th parade monitoring from his phone. This is real-world use: hand off the churn to a system that scales, you keep control of direction.

Want to see how it works: the complete process is public in the shared transcript and GitHub pull request. Willison doesn’t hide limitations or failures—it’s the best example of how you actually work with a coding agent on a real project.