Tell Claude's new model "never" and it won't budge. Entropy tells what to write instead.
Claude Fable 5.2 treats any instruction with "never" as absolute. Its asking guide suggests "Ask permission before deleting files" over "Never delete files" when you still want to say something.
Newsletter writer Ruben Hasid turned the guide into lines that anyone can paste:
1. Find "always" and "never" in the instructions and in your memory Claude. In gray areas, it says "ask me first".
2. Don't just say "go". Say what you approve: “Yes, rewrite the 3 welcome emails. Leave the invoice template alone. ".
3. Should finish: Finish all work before quitting. Do not suggest to continue: Continue. Ask me before you delete something. 4. Ask for a report you can read: finish with a summary for those who haven't seen any of the work: key points first, simple words.
In Anthrop's replay of real coding sessions, the longer end-of-work text roughly halved the runs that terminated with work remaining.