The main thing here would need standardisation is the environment in which the skill operates. The skill instructions are interpreted by the AI, any support scripts are. Interpreted by the environment.

You don't want to give an English description of how to compress LZMA and then let the AI do it token by token. Although that would be a pretty good arduous methodical benchmark task for an AI.