Design and optimize prompts

Why it matters

A prompt change that saves tokens is only safe to ship when its quality is measured against a baseline that was recorded before the change.

What to practise

Provision a Microsoft Foundry project with azd up, deploy the Trail Guide agent with the v3 production prompt, and run the batch script over five fixed test prompts. Score each response by hand for intent resolution, relevance, and groundedness in an evaluation.csv, with a comment. Then repeat the run on Git experiment branches: a token-optimized v4 prompt and, if your region offers one, a lower-cost model. Compare the three CSV files and the token counts, and merge the winning branch. Watch which test prompts change score, not only the totals.

How it fits

The hand-scored CSV is the reviewer evidence 3.1.1 asks you to make usable: a template, a fixed sample, and a written reason per score. Comparing v4 with the recorded v3 baseline is the candidate-against-baseline decision in 3.1.3, made here by people rather than by evaluators.