They've exposed the system prompt via the network requests. I'll give them the benefit of the doubt and assume that was intentional, and not because the app was blindly vibe coded ;-).
Here's the prompt as of around release-time yesterday. You can verify in your own network requests.
For those of us who are implementing similar AI-products, it's an interesting opportunity to see what worked for someone else. While it may be vibe-coded, it also... works.
In this case, I was interested in how they thought about the design aesthetics, and the prompt led me to the specific "skill", which the agent was happy to spit out for me when asked...
The Claude Design tool exposes its system prompt via the API requests it makes. I'll give them the benefit of the doubt and assume that was intentional, and not because the app was blindly vibe coded ;-).
I've linked the prompt as of around release-time yesterday. You can verify in your own network requests.
Thanks! No personalization on taste yet. "Personally", I'm not saving items that I necessarily like more anyways, but rather ones that I want to check out later. Future work to add an "upvote" button that drives personal recommendations on taste, though!
Hi! I created an "Escape Site". It has a score of challenging puzzles designed to make you think outside the box, and think differently about the way websites work. The success rate is about 10%. How do you fare?
Here's the prompt as of around release-time yesterday. You can verify in your own network requests.
https://gist.github.com/mcrowe/9da081f52e0740886d39e730852db...
reply