a YC funded AI support agent company open sourced their prompts used in production

lots of details to learn from

overall their prompts read less like text and more like programming

https://parahelp.com/blog/prompt-design (bruh that link preview image)

you might like https://github.com/galligan/mixdown


Or my workflow file that runs in cursor (as a workflow) https://github.com/alexpaden/identity-ai/blob/main/workflow.md

Mixdown is cool because xml costs a lot of tokens standalone but I’ve tested these things out before and otherwise lost non-formatted prompt data is recognized by llms when wrapped in xml

Verbose is probably not the right word, less wordy*

Is this how the user is expected to prompt too? I wonder how it affects accurately being less verbose

Very interesting ! Much less ambiguous than natural language 😍