Prompts Drift. Skills Don't.
Over the past week, my Claude Code setup started feeling off. Something was drifting, and after some digging, I tracked it down to three culprits. 1. The mystery .claude/memory folder. Anthropic seems to have quietly added a memory folder under .claude, and it was busy saving every bit of feedback I gave Riggs (my AI) about doing things differently. Problem is, Riggs and I already had our own rules for updating system prompt files — and that new memory folder was effectively letting him sidestep them. 2. Hidden log folders I didn't realize I'd green-lit. During a brainstorm, I'd casually approved a plan that spun up log folders under .claude. Because they're hidden, they never showed up in my VS Code file tree. The result: three separate locations holding three different sets of documents, all quietly diverging. 3. The prompts themselves. This was the real kicker. My prompts leaned on the folder structure and left room for interpretation. That wiggle room was the final straw that tipped the whole system into feeling wonky. Prompting isn't bad on its own — but for routine work, it gives Riggs too much room to improvise mid-task. My fix: stop prompting for routine processes and start turning them into Skills tied to custom commands. No more ambiguity from near-identical phrasings, hardcoded instructions every time, and far fewer runaway files and folders piling up in the dark.