Comment by ra

The best advice is to use evals to improve any code that produces output that goes into the LLMs context.