Rules files may often contain context important for the LLM to have during it’s thinking/planning. But for both Claude and Gemini, I’m consistently seeing the rules pull in after the thinking/planning has occurred. This is limiting how well the thinking models incorporate the rules files.
For Agent Requested rules, it seems to pull rules randomly at some point after the thinking. The image below shows an example where the LLM already thought, came up with a plan, then pulled the rules file
For Auto Attached, it seems to pull rules when the first relevant file is reviewed. Unless a relevant file was manually added to context, this happens after thinking.