You give the model a workspace and rules:

1. Inspect the corpus


2. Decompose the problem


3. Solve sub-parts independently


4. Verify logic and assumptions


5. Commit only when confidence is high

This feels less like chatting and more like working with a junior analyst who shows their work.