What they did
Montesinos maintains content for touch-screen directories in shopping centres and used three scripts that refetched and rebuilt everything on every run. In September 2025 he described to ChatGPT, running GPT-5, a reusable replacement: fetch shop data from the web into YAML, skip unchanged sources, declare dependencies, rebuild only what changed, fetch in parallel and report changes. He expected a design discussion. Instead ChatGPT returned a ZIP with a working skeleton, and a second ZIP with hierarchical fetching when he raised that need. He committed the second ZIP the next morning. GitHub Copilot then helped him tidy, extend and release it, and in November 2026 Claude Code compared the ZIP with the first commit and re-ran the original code against a small test site.
How it went
The committed Python matched the ZIP line for line: 527 lines in nine files. It needed one fix, because it called a Typer function that does not exist. Later cleanup with Copilot added 1,642 lines on day one and introduced two bugs of its own: plugins that were never imported, so runs silently did nothing while checks passed, and a packaging config that left the plugins out of the wheel. Both were fixed by version 1.0.0 on New Year's Eve 2025. He says it has run flawlessly for nearly a year, checking a mall site several times a day and triggering alert emails when tenants change. A planned rules engine and parallel builds were never implemented.
Worth knowing
The generated code worked; the human-directed refactoring is what broke it, and tests did not catch the regression. The article itself was written by ChatGPT from conversations with him, with a closing note in his own words. The code is open source under MIT.