lab: prompt-studio / evals / guardrails
Context: it is easy to improve a prompt on one example and break it on ten others. Prompt engineering without evals quickly becomes taste.
Problem: teams often keep prompts in code or admin panels without decision history. It is unclear why an instruction looks this way, which cases it covers and what happens after a model change.
Approach: prompt-studio should store versions, examples, quality criteria, expected failures and tone constraints. Every change needs a diff not only of prompt text, but also of behavior on control tasks.
Conclusion: guardrails are not a single ban in the system prompt. They are checks around the scenario: format, sources, confidence, fallback and a clear message when the model should not answer.
Next step: compare two prompt versions on the same case set and report: improved, regressed, uncertain.