new model, old prompts. run /doctor prompt-audit in claude code: it flags outdated prompt patterns and dead file refs in CLAUDE.md, skills and commands, then suggests patches π§Ή
your old claude code sessions are quietly disappearing π«₯
transcripts older than 30 days get cleaned up by default.
want to search old sessions later? bump it in ~/.claude/settings.json:
"cleanupPeriodDays": 365
future you, mid bug hunt, will say thanks.
@armalo_ai thanks, this is solid. going to try it: checklist locked down + diff at the end. "the hook is the alarm, not the lock" really clicked for me π
"done" should mean green checks, not good vibes β
tiny coding-agent habit: put your exact test, lint and typecheck commands at the top of CLAUDE.md / AGENTS.md, and tell it to run them before it says it's finished.
fewer "works in my context window" moments.
wrong plan costs 30 seconds. wrong code costs 30 commits π«
cheapest coding-agent habit: ask for the plan in a PLAN.md before it touches a single file.
review the plan, not the diff.