Claude code kept breaking my projects until i changed how i used it
shipped two ios apps this year using claude code with zero swift experience. trot, a couch to 5k thing with apple watch. hoist, an invoicing app for builders. third one threadsmith is in build now. completely different categories, same system every time.
first few attempts were chaos. ask for a feature, get code, paste it, something breaks, ask claude to fix it, breaks something else. three hours in i had a project nobody could reason about, not me, not claude.
what fixed it was treating claude code like a junior dev who needs a spec, not a chatbot.
lock every architecture decision before writing any code. write 3 to 5 non negotiable rules, the kind of thing that compiles silently when wrong but blows up later. break the build into phases, one phase per session, finish before moving on. storekit last, always.
the unlock was realising claude code doesn’t need to be smart, it needs context. give it a locked spec and it writes consistent code every time. don’t, and it pattern matches and contradicts itself. wrote the whole methodology up as a build system. seven docs, templates, the lot. happy to talk through the approach in the comments if anyone wants. app store links to trot and hoist on request too.