FitLog dev swaps Claude Code prompt warnings for a 48-line CLAUDE.md harness

The developer's starting point was a pattern of repeated failures: Claude Code would open .env files, try to edit ios/ and android/ directories managed by Expo, or modify originals in components/ui/ generated by Reusables. Adding "please don't" lines to prompts only made the instructions longer, and compliance felt like a matter of luck. The reframing was to separate rules that cause accidents if broken from recommendations that can be fixed afterward, and to enforce only the former mechanically.

The resulting structure splits labor between documentation and hooks. CLAUDE.md was cut to a 48-line map pointing to six docs files — architecture, db-schema, reusables-guide, conventions, harness, and design-docs — so details are loaded only when needed. The four hooks include protect-env and protect-native, plus a formatting hook that skips silently if Prettier is absent and a Stop-time summary of git diff --stat. Five Skills cover /commit, /review, /deploy, /cleanup, and /new-screen, with the high-impact /deploy marked disable-model-invocation so only an explicit user call can trigger it.

The write-up also names its own limits. The developer could not read the original OpenAI article because of a 403 error and designed from summaries instead. A broken README with typos and missing sentences let Claude fill in guessed meanings, showing that a corrupted map defeats the harness. And while the hooks passed a 10-case sample test, real-session behavior is still to be confirmed.