Few-shot examples come from the eval set, not from imagination
Examples you write yourself are clean and typical. Examples from production are messy, and they're the cases the model actually gets wrong. From the eval set's failures, pick the three that a good example would have fixed.
input: <a real ticket, trimmed>
output: billingSwap them out when the failures change. An example that no longer matches a failure mode is just tokens.
ai-engineeringllm