Prompts that actually work

Eight states sued OpenAI this week alleging ChatGPT was marketed as safe while its risks were withheld. Whatever the outcome, the practical lesson for anyone buying AI is that a vendor's safety page is marketing copy until you test it. These five prompts turn a vendor's own claims into questions you can check — run them against the vendor's documentation, not against the vendor's chatbot.

Fish Audio's Drama 3 landed this week at 93.2% on a pronunciation robustness benchmark — which tells you about pronunciation robustness and nothing else. Benchmark scores do not predict whether a voice works for your script, your languages or your budget. These five prompts build the test set a leaderboard cannot give you.