Hi all!
An idle question:
It hit me today that “prompt engineering” points at the lack of a structured language for how text works as an input into SD, whose affordances are discoverable.
One step is just to experiment with the tokenizer (like, could a language take interpolations in the input string, like “a painting of a {cat:.8, dog:.2} in the style of Picasso”
I just wonder if anyone has seen any research on say, “structured prompt languages,” or a kind of “PromptQL.” It occurs to me like databases have SQL, surely there are better ways of syntactically organizing prompts into SD.
I actually just created a prompting language call Convo. They syntax allows you to write both your prompts and define functions for LLMs to call. The functions are fully defined in convo but can also call out to javascript functions. You can also use the vscode extension to execute your prompts directly in vscode
Here is a link to the CLI and you can find the vscode extension by search for “convo” in the extensions tab.