Keep Prompts Under Review

A small wording change can make an agent more helpful, more evasive, more verbose, or more expensive. Prompts are not configuration — they are behavioral code. They deserve the same safeguards.

• Clarity. Can a colleague understand the intent without asking you?
• Scope. Does the prompt stay inside the task boundary, or does it invite scope creep?
• Safety. Does it ask the model to reason about private data or make decisions it should not?
• Cost. Does it include unnecessary context or force long outputs?
• Examples. Are the few-shot examples still representative and correct?

1. Treat prompt changes like code changes: open a PR, explain the goal, and show before/after behavior.
2. Run a small test set against the new prompt and compare outputs.
3. Look for regressions in unrelated tasks that share the same base prompt.
4. Roll back quickly if behavior degrades in production.