Tl;dr: Yass leverages RFC2119 keywords in a minimal Yaml format, relying on trained in behaviors for those keywords to provide strong model behaviors.
The yass CLI itself is all that you need - simply tell an Agent to look at the shakefu/yass repo README and they get all the information needed to proceed.
This was validated with hundreds of experiments and trials against different models and platforms. The most significant of which was a single Yass spec tree reproduced byte-exact output Yass CLI implementations across 7 different languages (TypeScript, Python, Go, C, C++, Erlang and Elixir) in sandboxed Claude Code harnesses using Opus 4.6, rather than expensive frontiers.
Give it a try, open GH issues if you find things that aren’t working (or make your agent open them), or if you have feature requests.
Enjoy!
bilbo-b-baggins•41m ago
A yass spec tree with no CLI downloaded was used in the sandbox testing to validate the specs alone is enough to drive strong model behaviors and reproducible implementation output.
The CLI acts as a token saver and steering context letting the agent pull spec references without needing to consume the entire tree.