Add Claude skill for unity mcp

[dsarno]

Things like:

• use refresh_unity tool after writing scripts (including new tests) to make sure they compile before running them -- and read_console to check for errors.
• use batch_execute for batch execution of parallel tools

And then some schema tutorializing for the shape of paramters and types when making tool and resource calls.

Anyone think of any others? @Scriptwonder @msanatan and community?

[Scriptwonder]

What I can think of:

1. Use capture_screenshot to check if the visuals make sense and use console_log(cannot remember the detail name) to check compilation errors after each change.

I think at the start of the skill we should just generally introduce all tools' functionality to the LLM, and then introduce some common practices like you mentioned here. I can comment more when I think of it.

[msanatan]

Maybe a more general guidance or reading resources before the tools? To me that's wha the SOTA models do so much better than cheaper ones, they list resources and read the right ones

[Scriptwonder]

I just used Opus 4.5 Deep Thinking to co-create a skill at #659 as a starting point, incorporating all our feedback. Feel free to edit anything!

PS: I feel evaluating the performance with/without skills is very tricky based on my recent experience doing exactly this kind of evaluation, so I put it as a to-do.

[dsarno]

@Scriptwonder Thanks, this is awesome! I'm sure we'll learn more about how effective / efficient skills are as time goes on -- but excited to see if this will help LLMs avoid some common mistakes with the mcp.