I’m hitting a recurring issue in long sessions where the model starts drifting away from the core instructions after the context gets too deep (the classic "lost in the middle" problem). Even if I set strict rules in the System Prompt, once the history gets long, the model starts ignoring naming conventions and formatting—especially when I'm working with complex CSS/DSL structures.

​Right now, my only fix is to manually paste my "Remember to do X, Y, and Z" instructions at the end of every single message, which is pretty tedious.

​Since the newer models (Claude 4.5, etc.) have such a strong recency bias, being able to automatically inject a "Final Anchor" at the very end of the prompt would be a huge UX win. It would keep the model's output aligned without needing constant manual reminders.

​Is this something that would be easy to implement? Cheers!