experimental: synthetic instructions

[spoons-and-mirrors]

Why

I've noticed that adding another snippet in the system prompt about how to use a specific tool help the llm adhere to the instructions more closely.
I've also noticed that if you move that snippet to the user message level directly, the model is even more likely to do what you want.

What

The synthetic instruction is a feature intended to do just that, inject a synthetic text snippet at the user message level. Particularly useful for experimental and plugin tools. This leverages recency bias in models to better adhere to tool instructions

PS; this PR is very much slop

[Tarquinen]

This is awesome lets get this in!