Improve a prompt from graded test cases

Improve this prompt using graded test cases.

Work in this order and *do not skip to the rewrite.*

1. **Rubric first.** What does a good output have? Three to five checkable criteria, each pass/fail, not a score out of ten.
2. **Test cases.** Six to eight inputs: two typical, two edge, two adversarial, plus the failures I described. State what each is testing.
3. **Predict.** For each case, which criterion the current prompt will fail and why. Base this on the prompt's actual wording, quoting the line responsible.
4. **Diagnose.** For each predicted failure, name the exact clause at fault — missing constraint, ambiguous instruction, conflicting rules, or no fallback for missing input.
5. **Rewrite** — changing only what the diagnosis calls for. List each change with the case it addresses.
6. **Re-predict** the same cases against the new version, and say which failures should now pass and which remain.

Rules:
- *Do not rewrite anything the tests did not implicate.* Untargeted rewrites are how a working prompt regresses.
- If a failure cannot be fixed by wording — it needs information the user has not supplied — say so. That is a task design problem, not a prompt problem.
- Keep the rewritten prompt no longer than necessary. Length is where instructions start contradicting each other.