Claude Code ran a solo business for 30 days: 14 tasks, 2 deals

The record is unusual less for the tools than for the accounting. The operator deliberately measured something other than sales: how long the AI could run while the human slept, and how often the human-only buttons stayed human-only. On those two measures, the setup held by day 30, even as revenue stayed at zero and a week-five audit forced a rebuild of the selling side.

The incident log explains why so many rules exist. A duplicate video upload nearly happened because completion was judged from the text of the board rather than a dated completion record. Two Claude Code agents sharing one git worktree led the reviewer to run git restore and wipe uncommitted work. Fake test spreadsheets passed while dropping leading zeros in phone numbers and seconds in timestamps. Each failure turned into a rule such as distinguishing empty from zero, forbidding git add -A, and stopping without input when a screen does not match assumptions.

The pitch data also redirected the strategy. On day 22, reviewing 151 pitches showed the median application count at send time was 94 and prices clustered near budget ceilings. The two wins came from listings that were fresh, had around 10 applicants, and had budgets under 10,000 yen. That group returned 8.8% replies, while the crowded, large-budget group that made up 64% of sends returned 0 of 81.