I usually talk to LLMs as if they were a child, gradually interacting with them and introducing them to the code.

Then, once it's familiar with the project's code, I ask: How would the LLM advise me to conduct an audit? How many parts should I break the audit down into to better review the code, and to ensure it doesn't prioritize saving money or tokens, because accuracy is more important than money.

The LLM breaks the audit down into several parts for me. For example, right now I'm auditing Monfluo (a cool Monero wallet).

The LLM (Kimi K3) split the audit into 9 parts, and after it finishes Part_1, I then ask it to double-check the logic it documented in a separate file, and it re-examines what it found during the first check.

I also tell the LLM to create its own .md memory file, to refer to that memory file whenever it needs to document something (make changes), and to always refer to the memory file whenever it needs to recall something. This memory file is also useful across different sessions.

After completing the audit using the LLM, I manually verify the findings to double-check that the bug actually exists.

And yes, Kimi K3 is impressive.