I've recently become intrigued by representing concepts so that AI can co-contribute effectively and token-efficiently (typically: find a good way to represent something as semi-structured sequential text), but also without compromising the human lens on the representation. We shouldn't accept a downgrade of the human knowledge representation experience just to make it AI-accessible. That's especially true if traditionally non-dev personas need to contribute, and they almost certainly find "weird text format + git" much worse than their current authoring/viz tools.

I'm excited to see how standards for semantically representing different kinds of knowledge emerge in the next few years!

Successful examples I can think of to mix in are open standards like DBML for schemas/E-R, LikeC4 for architecture, diagrams-as-code ideas like Mermaid, all of which LLMs seem to "get" well (or can be told about from a short EBNF prompt). Crucially, they also have pretty human viz forms, and you can you can just ```code block``` inline them in Markdown next to natural language. And you can get LLMs to help you author the syntax.

Harder to crack is stuff where there's implicit human meaning in spatial layout and colour, like in complex spreadsheets or Miro. I haven't found good alternatives for those yet.

My own attempt in my (data engineering) domain is https://equalexperts.github.io/satsuma-lang/ for AI-and-human source-to-target mappings and transforms. A succinct structured text representation that allows natural language, but also nice viz and LSP/grammar tooling that helps agents not to have to slice and dice big docs token-inefficiently to reason about things like lineage or completeness or undefined sources.

At first glance, it's just a nice version-controllable, parseable DSL. But I also made succinct prompts with the grammar that lets LLMs produce and reason about Satsuma, a language server, CLI tools for the AI tools to use to navigate specs token-efficiently, reason about lineage, pretty viz in vscode plugins/syntax highlighting, some agent skills etc. There are metadata conventions for succinctly representing a lot of quirky formats and capturing common analytics conventions (scd2/Kimball/datavault/etc c.)

Yes, yes, I may have gotten carried away.

But I'm finding it really useful as a specification tool in projects for reverse engineering mappings from code/workflows, generating new code (dbt, Spark etc.)

This definitely isn't something I would've had the bandwidth to push this far before AI!