At this point, it really does look like the key to keeping LLMs on the rails in your dev workflow is:
* Small feedback cycles, solving one problem at a time
* Prompting with test cases/usage examples
* Continuous testing
* Continuous code review & refactoring
* Continuous integration
* Effective separation of concerns in the design
LOL.
What you've got here is essentially the engineering practices of Extreme Programming.
It could even be argued this is a statistical case for their efficacy.