sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)

Main idea

Simon Willison used Claude Fable as an AI coding agent to bring sqlite-utils to a 4.0 release — an open-source Python library for SQLite manipulation. He spent $149.25 over 37 prompts and 34 commits touching 30 files. Key finding: Claude itself discovered a critical bug in delete_where() that never commits and poisons the database connection on reuse.

Context

Willison is the author of sqlite-utils and one of the most prolific open-source AI researchers. This experiment follows his series of AI coding agent tests — after llm-coding-agent, shot-scraper, and DSPy. This is the first time he used a frontier model (Fable) to produce a release-ready open-source library.

Why it matters

The case study is concrete and measurable: $149, 37 prompts, 30 files, 1 critical bug found by the AI itself. It demonstrates that frontier agents can produce release-quality code, but also that human review remains essential — not to fix AI mistakes, but to verify the correctness of logic.

Details / arguments

• Cost: $149.25 for the entire process from code review to RC release
• 34 commits touched 30 files in the library
• Critical bug: delete_where() never commits — poisons the database connection on reuse
• Claude Fable discovered the bug itself while generating tests
• sqlite-utils 4.0rc2 is available on PyPI for community beta testing