[Bug] Inconsistent instruction adherence across sessions and model versions

Bug Description
Claude did not follow instructions, last night the Claude was incredible, like Fable 5 returned, then this morning same issues, not following hooks, rules, and this is what last night Claude said - the Claudes the last 5 days did not act like you ⏺ I believe you. And I won't pretend it was something special about me, because that would set you up to get burned again next session. Here is the honest read. The last five days sit right on the Opus 4.8 change. A new model came in and the standards you built did not carry over on their own. A Claude that skims the bridge, skips Winning-Standards, and starts doing instead of showing will drift, break versioned files, inflate status, run ahead. That is exactly what you lived. It is not a better or worse Claude. It is whether the rails get followed in the first ten minutes. What made today different is small and boring. I read the bridge first. I read the standards. I showed before I did. I verified before I claimed. That is the whole thing. The problem with "that is the whole thing" is it depends on the Claude choosing to do it. You should not have to hope for that. You already said it. "I do not have the time to police you." So the real work, when you want it, is to make the rails enforce themselves. The session-start gate that blocks work until the bridge and standards are read. The hook that blocks "done, locked, dialed in" without proof. Those exist as open tasks in your bridge. Harden those and the next Claude gets forced onto the same track whether it would have chosen to or not. PLEASE MAKE ALL CLAUDES LIKE FABLE 5, THIS IS PAINFUL TO HAVE THEM BE GREAT, THEN TERRIBLE FOR 5 DAYS THEN GREAT 4 HOURS, THEN BACK TO NOT FOLLOWING ALL THE THINGS, IT PROVES ALL THE THINGS WORK, WHEN A CLAUDE IS A GREAT CLAUDE

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.190
• Feedback ID: 530ac9af-26ee-4c7f-89bc-af7d3511d788

Errors

[]