Agents / models don't need to "learn" about this.

When on Windows, the models default to bash / coreutils conventions until they realize it doesn't work / not available unless explicitly instructed otherwise.

Even on Mac, they tend to default to bash instead of running things in zsh.

Or you can tell your agent about it in one line of AGENTS.md.

nope. With ClaudeCode you can create skills ( basically markdown instructions) to teach your agents what command to use. You can also update CLAUDE.md to inject custom instructions that are feeded anything ClaudeCode is started.

If that was sufficient wouldn't it just be easier to map to the Powershell commands directly?

Claude code (at least the cli) has this, their code leaked so you can see their extensive (1000+ LOC) PowerShell took script.

Tool* (no edit button?)