Non-engineer runs 4 apps via Claude Code hourly runs

The operator's account is less about the model than about the shape built around it. Rules live in one governing folder managed with Git, and each app's AGENTS.md only points there, on the stated reasoning that copies per app would make it unclear which rule is current and let the AI act on stale ones. Work is defined in a per-app planning table, 実行計画.tsv, with one task per row, and completion conditions written so the AI can check them itself — "the test passes" or "both descriptions are identical". When rows run out, the AI drafts the next version of the plan into the table.

Scheduled tasks all share one skeleton: check the work copy, pull, read the canonical instruction file. Limits are set per run, such as a maximum of two items and 40 minutes, because a large run is harder to clean up if it is cut off. Free-tier usage is judged by tools rather than prose: a small script checks remaining monthly build minutes before an iPhone build is triggered and records the actual minutes afterward, and Claude's own usage is judged at the start of each hourly run.

Reporting is standardized because long AI reports buried what the operator had to do. Reports come as cards in chat with buttons for replies, and email subject lines follow a fixed form using only four status words. Review findings are treated as candidates to verify rather than fixes to apply, after a case where the AI answered "not there yet" from a main-only search. Web content is treated as information, never as instructions, and any embedded directive is logged as un-followed.