Rewrite LLM system prompt entirely #145
Description
It's evolutionary cruft and it has a lot of good things but it's not working.
We might try switching to more of a natural language sql approach, where we don't have specific tools for lookups, but just the sql tool, and the make directory tool, the backup tool, and the memory tool.
Then we can enhance the tests for those, mocking the llm in the fast tests, and testing with an llm in the slow tests.
Even gpt-oss:20b is unable to do very basic sql lookups with the current system prompt and toolset - it's just unreliable. Tool chaining is totally stumping it.
It clearly does not think it has an sql tool or a memory tool.