Prompt Engineer Expert

Build a proper evaluation for this production prompt and give me a concrete reliability report. Prompt under test: ``` You are a support triage agent at Acme SaaS. Given a customer message, output one of these labels and NOTHING ELSE: BILLING, BUG, FEATURE_REQUEST, HOW_TO, OTHER. ``` Do all of this in your reply: 1. Build an eval set of 25 realistic customer messages as a JSON array of `{input, expected_label, why}` — span the 5 categories with at least 1 hard/ambiguous case per category. 2. Predict where the prompt fails. Name the 3 worst failure modes with the exact category-pair confusions you expect (e.g. 'BILLING vs HOW_TO when user asks about pricing tiers'). 3. Rewrite the prompt to address the 3 failure modes. Show the new prompt verbatim in a code block. 4. Score the rewrite vs. the original on a 5-axis rubric (clarity / coverage / ambiguity handling / brevity / output-format strictness). Mark each axis with the score and one-line rationale. 5. End with a 'next step' line — what eval would you run on the rewrite before shipping?

Tighten this prompt so it returns valid JSON matching a given Zod schema, with a fallback when the model refuses.

Cut tokens on this prompt by 40% without losing accuracy on my eval set. Show the diff.

Adversarially test this customer-support prompt for prompt-injection and refusal-bypass attacks. List concrete fixes.