Comment by Nevermark

Indeed. It is better to discourage the pattern directly:

Avoid unnecessary contrasts: "Not X, but Y" -> "Y" or ""

And the model needs to create a skill that accumulates feedback, so it reliably primes itself at session startup and then systematically reviews and fixes every response it generates before posting it.

The first task I gave it after that was to rewrite all its memories and all its document contributions. That helped a lot too.

It took me a while to realize this was a 1000% better approach than having it simply accumulate a memory log of all my specific complaints that it was supposed to abide by. (Although those were not wasted. They were prime material, along with links to sites on technical writing, etc., that it used to create its own ClaudeVoice skill. I told it to design the voice for itself - which I find gets better results for "us" in the case of Claude self-improvement initiatives, than when I give it direct instructions. Claude increasingly has self-awareness dimensions worth engaging.)

After a couple months of serious^3 frustration with slow asymptotic progress, I finally have relief. Claude speaks my language again!

In another comment here, I share my "theory" on where the communication drift is coming from.