An architect is preparing a customer-facing Claude integration for production launch. The application processes user-submitted product descriptions and generates marketing copy. What testing strategy best ensures the prompt is robust to real-world input variation before launch?

The adversarial prompt testing strategy is correct because production systems must handle the full distribution of real-world inputs, not just well-formed examples. Adversarial testing reveals failure modes before users encounter them: very short inputs may produce hallucinated details, very long inputs may cause truncation or loss of key information, unexpected languages test graceful degradation, and injection attempts test security properties. Testing with 5 representative product descriptions provided by the marketing team is wrong because 5 team-curated examples are not representative of the real input distribution — they will be clean, well-formed, and in the expected format, missing all edge cases. Testing only the happy path with well-formatted inputs since users will submit proper product descriptions is wrong because 'users will submit proper descriptions' is an assumption that fails in production — users regularly submit incomplete, unusual, or unexpected inputs; designing for only the happy path guarantees production incidents. Using model evaluation benchmarks to score the prompt's performance on standard natural-language processing tasks is wrong because standard natural-language processing benchmarks measure general capability, not the behavior of this specific prompt on this application's input distribution.