A rule that adds ceremony needs to get out of an agent's way, including a rule I wrote myself. It is easy to complain about a slow model while treating my configuration as sacred. If I asked it to request confirmation at every step, part of that queue of questions has my signature on it.
Before demanding another answer from the model, I look at how I organized the execution and ask whether the rules I created to organize models are holding newer models back.
Anthropic announced Opus 4.8 on May 28, 2026, along with dynamic workflows in Claude Code, a research preview in which Claude plans the work and runs hundreds of parallel subagents in a single session. A tool that plans its own route deserves instructions that allow that choice. Otherwise I buy new capability and tie it to the old script.
I understand how the pile grows. One mistake produces a rule. Another request needs an exception. Eventually the file says to investigate everything, execute quickly, and stop to confirm things already authorized. That is what I want to catch when I review instructions. A document can sound reasonable line by line and become impossible when read as a whole.
My first question is which mistake a rule prevents. I want to name that mistake and recognize it when it happens. If the rule protects necessary behavior, it stays. If it prescribes a command sequence without explaining why, I check how much of the agent's route I have decided for it.
Deleting everything would erase useful knowledge too. The review should be small enough to understand: one questionable instruction, the behavior worth preserving, and comparable executions.
The announced feature does not do this investigation for me, but it makes frozen instructions harder to defend while the tool underneath changes. I will review rules by the errors they prevent and rewrite the ones that only prescribe steps, so the agent gets limits that are easy to understand, a delivery I can check, and freedom to choose the route.