จัดเวอร์ชันพรอมต์และประเมินเอาต์พุต

A reworded prompt can change behavior overnight. So you treat each prompt as a versioned artifact, with history, review, and the ability to roll back. 📝

Hard-coded strings scattered in code are impossible to track. Keep prompts in a registry or files so every change is visible and diffable.

Give every prompt a clear version id you can pin in code. Then you always know exactly which wording produced a given output.

Collect real inputs with the answers you want. This eval set is your yardstick for judging whether a prompt change actually helped.

For tasks with a known answer, compare output to the expected text. A simple metric like exact match or F1 gives you a fast pass-fail signal.

For open-ended text, ask another model to grade the answer against a rubric. This LLM-as-judge pattern scales scoring beyond hand-written rules.

Judges can be biased or inconsistent. You keep them honest by spot-checking scores against a few human ratings on the same cases.

Before shipping a new prompt, run it across the whole eval set. You only promote it if the score holds or improves, never on a hunch.

A prompt that fixes one case can break another. Comparing per-case results catches these regressions before users ever see them.

Tools like promptfoo or OpenAI Evals run datasets, call models, and report scores for you. A framework turns ad-hoc testing into a repeatable suite.

Once a version wins on your evals, pin it in production and archive the rest. You now have a clear, auditable trail from prompt to result.

You need to score open-ended summaries with no single correct answer. What fits best?

You learned to version prompts, build an eval set, score with metrics or an LLM judge, and gate releases on results. No more guessing! 🎉