aichannel

💽 Let's Mixdown 💽

Just published Mixdown: an open-source multi-syntax markup spec that I've been putting together for my AI prompt engineering.

Feel free critique, copy, remix, or add issues & PRs if you're interested.

https://github.com/galligan/mixdown

More about Mixdown ↓

Made something, would love feedback!

It's a markup spec / ai prompting technique I've been using with some early success. I designed it to be straightforward to write, make use of variables, and produce consistent results.

Try it and let me know what you think!

Made something, would love feedback!

It's a markup spec / ai prompting technique I've been using with some early success. I designed it to be straightforward to write, make use of variables, and produce consistent results.

Try it and let me know what you think!

https://warpcast.com/mg/0x20b9cae5

Co-founder XMTP, building web3's secure messaging layer ➡️ https://xmtp.org

Decentralized open source consciousness | Values-aligned communities | citizen of Black Rock City

They probably would! Or at least a format that it might have recognized from code examples in its training set. 

Personally I'm a fan of constraints, so that was the exercise I tried for Mixdown. Limits the variables by which I can make tweaks and see if they work.

Interesting. I like it in general. Given these models aren't trained on this format, I wonder if just the structure (with any syntax) would perform as well

Yeah, it's definitely been something on my mind as well.

Could very well be that there's a sweet spot that Mixdown works for, and it might not be super complex, really long prompts.

Anecdotally, I took some time today to convert Repo Prompt's "Diff Protocol" prompt to see how it would fare, and the result was a 1:1 parity with the default prompt, written in XML.

been thinking more on the eval side of this and i'd be curious about the yaml tradeoffs in terms of size and ability to recall precisely from larger context, as well as instructions/workflow steps in and not in XML (prefer action steps), and more broadly the quality tradeoffs against size of tokens in context (my guess is xml is more exact and expensive)