💽 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 ↓

Mixdown is a syntax/technique for writing AI prompts that work well in most modern LLMs. In my experience it's been easier to create more complex prompts that return reliably consistent responses.

It mixes elements of XML (structure), YAML (data), and Markdown (prose, output formats).

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

The syntax allows for adding key:value pairs, which can then be used as variables within the prompt. This sets up a means to quickly swap out individual bits without reconstructing entire prompts.

Check out how GPT-4o handled this prompt (with no additional instructions):

The syntax allows for adding key:value pairs, which can then be used as variables within the prompt. This sets up a means to quickly swap out individual bits without reconstructing entire prompts.

Check out how GPT-4o handled this prompt (with no additional instructions): https://chatgpt.com/share/67e1da7f-ec10-8000-bc57-d68c19737c86

It's still (really) early days for putting this together, and while qualitatively it's working well for me in my experience, it's by no means proven to work yet in plenty of cases.

Ultimately if the only one that gets use of it is me, that's ok, as it's been fun and super helpful so far. But I'm hoping others get some value out of it too…

So that's why I'm open-sourcing it! Hoping to see what y'all think of it. Bring on the feedback!

/max • /bracket • @quickcast • /coys • maxp.co

Thanks! I really appreciate that.

It's really the culmination (and codifying) of a lot of things I've been trying in my prompts for a while now. The results for prompts compound when you just stick to a technique and try to make incremental improvements. 

Let me know if you try to use it!

really cool stuff! can tell the thought that went into this

Tech Lead at Rootstock Labs. Born in Colorado, based in Cyprus.

So far I've tested the prompts I've made with it across ChatGPT (4o, 4o-mini, 4.5, o1, o3 mini), Claude Sonnet (3.5, 3.7, 3.7 Extended Thinking), Gemini (2.0 Flash, 2.0 Flash Thinking), and Groq 3. So far the only one that had some inconsistency was Gemini 2.0 Flash.

But a little prompt tweaking, and it did fine!

This looks interesting. I usually use a combination of an intro paragraph with xml like brackets for specific parts. The structure here looks a lot better. Which LLM’s have you tested this approach on?

Beauty flashes have always been there, Capturing them along the feeling they bring.

bento.me/eyesobscura

Appreciate it! Admittedly, it was mostly an exercise in standardizing myself 🤣 …trying too many different approaches led to never being too sure what was actually working, and what was just generally AI magic.

cma.xyz. Founder /paragraph, ceo /mirror, previously at Google & Coinbase 

Rad! Let me know if you try it out with some prompts. I'd be happy to think through some prompt examples with you to see what works best. Would be fun to try out.

This is awesome. I’ve found myself struggling with prompt duplication and context management myself

this is cool! i've been messing around with markdown files for prompt engineering. will give this a try!