More generally I think testing AI by using its web search, code execution and ensembling is the missing ingredient to increased usage. We need to define the opposite of AI work - what validates it. This is hard, but once done you can trust the system and it becomes cheaper to change.
The issue isnt the LLM, its that verification is actually the hard part. In any case, its typically called “evals” and you can probably craft a test harness to evaluate these if you think about it hard enough
Not to mention the advantages it would present for iteration and improvement.
Just the format would be. There's no rigid structure that gets any preferrential treatment by the LLM, even if it did accept. In the end it's just instructions that are no different in any way from the prompt text.
And nothing stops you from making a "parameterized and normalized to some agreed-upon structure" and passing it directly to the LLM as skills content, or parsing it and dumping it as skills regular text content.