Best practice design decisions for AI-friendly error messages

[andig]

By the way, we already do what you're doing: https://github.com/modelcontextprotocol/go-sdk/blob/main/mcp/tool.go#L121.

@jba coming back to this out of curiosity. Working with https://github.com/jedisct1/openapi-mcp I've noticed that great effort was put into returning LLM-readable natural language responses even for errors. What's the strategy in this repo? Is there a documented approach? Are there insights that this may not be needed with the current LLM capabilities?

Originally posted by @andig in #268

[findleyr]

Any input from the community would be helpful here.