LLMs don't reliably follow instructions

Building your MCP server? Struggling with speed or token window size?

There's even another catch, LLMs don't reliably follow instructions!

There are some best practices for building tools as documented by Anthropic

"Provide extremely detailed descriptions. This is by far the most important factor in tool performance. Your descriptions should explain every detail about the tool"

What they don't document is sometimes that this doesn't matter because the LLM won't follow your instructions

It just won't do what you say!

For more: these ideas (as well as being based on my own experience) are discussed in Episode 51 of the excellent "Vanishing Gradients" podcast