- Author
- Simon Willison
- Published
- 18 December 2025
- Kind
- Personal blog post, argument with practical advice
- Original
- simonwillison.net/2025/Dec/18/code-proven-to-work/
- Subject
- Coding with agents
What it is
Simon Willison writes a long-running weblog about software and AI. This post responds to a pattern the author sees. Engineers submit large untested pull requests made with LLM tools. The post sets out what proof of working code means.
What it says
- The post says untested giant PRs left to code review are a dereliction of duty.
- The post says engineers must deliver working code with proof, in two required steps.
- Manual testing comes first: code you have not seen work does not work.
- The second step is an automated test that fails if the change is reverted.
- The post says coding agents should also prove changes, by manual and automated tests.
- The post says a computer can never be held accountable, so the human provides accountability.
What evidence it gives
The evidence is the author's own experience and habits. The post links one example of terminal commands pasted into a code review. It gives no numbers or comparisons.
Interests and limits
The post is opinion from one experienced developer. It describes the author's own workflow, such as teaching Claude Code to run CLI tools. It does not say how reviewers should treat contributors who skip the standard. The page also carries sponsor and newsletter notices.