Claude products actively nerf the model performance with 50k token system prompts. This is why I stick with Cursor. Your system prompts should be <5k tokens, preferably <1k. Cut out all the other fluff until you are left with the 5-10% most crucial tokens.