what should i use to benchmark system prompt changes

Use a small, repeatable eval suite rather than ad hoc testing.

Best options

1. LLM eval harness / test set

   • Create 20–100 representative prompts your system prompt is supposed to handle.
   • Include expected behaviors, edge cases, and failure cases.
   • Run the old vs. new system prompt on the same set and compare outputs.

2. Human rating rubric

   • Score outputs on a few dimensions like:
     • instruction following
     • correctness
     • tone/format compliance
     • refusal/safety behavior
     • hallucination rate
   • Use the same rubric for every change.

3. Pairwise A/B comparison

   • Show outputs from prompt A and prompt B side by side.
   • Prefer this over absolute scoring when subjective quality matters.

4. Automated checks

   • If you care about structure, use rule-based validation:
     • JSON parses
     • required fields present
     • banned phrases absent
     • length constraints met

5. Production-like replay

   • Re-run a sample of real user requests against both prompts.
   • This is often the most useful benchmark for system prompt changes.

What to measure

• task success rate
• format compliance
• refusal accuracy
• consistency/stability
• latency and token usage
• regression rate on important edge cases

Practical recommendation

If you want one setup:

• build a fixed eval set
• run A/B comparisons
• score with a short rubric
• add automated validators for anything structured

If you want, I can help you design a benchmark template for your specific system prompt.