[simonw]

I'm pretty confident in it: I saw the same result from a few different attempts, it's consistent with how the app itself appears to work, and the GitHub Next Twitter account tweeted my article without noting any inaccuracies.

Generally my past experience has been that leaked system prompts aren't hallucinated. Models are very good at repeating text that's just been fed into them.

[kelsecon]

Hi! Kelsey here (PM from Spark team). I can confirm the system prompt shared in Simon's post wasn't hallucinated. We worked hard to instill good design principles into the prompt. We put it through its paces with a number of designers that are some of the best in the industry. I also went to Parsons School for Design and can confirm this is 100% something they teach you at a good design school! ^-^