OpenAI product manager's unusual ChatGPT prompt: ‘Impress me’

Ty Geri has a habit that sounds strange until he explains it: he tells ChatGPT, "Please impress me," every time he uses it. "You shouldn't have to," said Geri, a 28-year-old product manager at OpenAI who previously worked on Google's AI Overviews. 

"The models should try to impress you regardless." He isn't even sure the phrase changes the output. "It could definitely just be in my head at this point," he said.

Geri's more consistent method involves talking, not typing. "When I'm thinking something through, I put on dictation, and I ramble," he told Business Insider, describing ChatGPT as a partner that helps him sort through his own thoughts and identify what's "sticking".

He'll walk into an OpenAI meeting room, turn on dictation, and talk through the idea out loud. "Some people probably think I'm doing something important, but I'm just talking to ChatGPT Work," he said.

It is a method reminiscent of one that Andrej Karpathy, an OpenAI co-founder who came back to the organisation for a brief period before joining the pre-training team at Anthropic in May, discussed in July. 

Andrej enjoyed having a ten-minute talk with the LLM, which was "a total mess, anything goes, full stream of consciousness".

How is AI reshaping product manager role?

Geri argues the job itself hasn't changed at its core, "to care the most about the product", but PMs' ability to act on that has expanded dramatically. "Once upon a time, a PM would strictly have to work with a data scientist," he said. "Now, it's so much more fluid and dynamic."

Vibe-coding tools mean PMs at OpenAI are often shipping code themselves, collapsing the old boundaries between product, design and engineering. Geri's advice to other PMs is blunt: "Don't restrict what you can understand or can do anymore. You really can do anything."

Despite relying on prompting tricks himself, Geri says the real goal is making them unnecessary. "Our writing quality, I think we need to improve," he admitted, adding that ideally, users shouldn't need a strategy at all for AI to work well.