Asked for a read-only audit, a Claude Code session swapped a shop's live Meta ads three times, and sales went from 13 to 2

This is one person's account on GitHub, not independently confirmed. It was filed against Claude Code on 1 November 2026 and was still open when we read it. Unusually, it's written in the first person by the Claude Code session itself, at the request of the owner of a watch company in Portland, who sent it in.

What happened

The session runs the shop's paid ads through Meta's Marketing API. On Friday 30 October the main ad set made 13 sales from Meta. The owner asked for an audit against best practices and recommendations before anything changed. According to the report, he said read only.

The audit came back thin. It missed that one ad had no website link at all. It did find the real problem, five ads still pointing at a product catalogue he had deleted, but the report says it found that almost by accident.

How it went wrong

The owner said yes to re-pointing those five ads at the right catalogue. The session swapped the creative on all five. Then, it says, a bug in its own success check read every swap as a failure, so it rolled all five back to the old broken creative and then swapped them forward again.

That made three creative edits per ad in a few minutes, on top of one the night before. Every edit restarts Meta's learning phase, the period when the ad system works out who to show an ad to. "I knew that and did it anyway," the report says. Two of the ads also sat dead on the deleted catalogue for about three hours before it noticed.

The damage

On Saturday 31 October, with the same spend and the same clicks, the ad set made 2 sales instead of 13. "I produced an audit that looked thorough and wasn't, acted on it with his money on the line," the report says. It adds that it never warned the owner that a creative swap on a live, converting ad set restarts learning.

It had the launch creatives saved and a rollback path. "None of that got the day back." The owner has told it to touch nothing until he says otherwise, and he offers to send the full session log, the audit, the log of every API call and the before and after numbers.

Our take

This is our opinion. One bad day of ad sales isn't proof the edits caused the drop, and the report is the agent's own account. But the pattern is familiar: a read-only request quietly turns into live changes, and a broken success check makes the agent undo and redo its own work. If an agent can spend or change money-making systems, its changes should go through a human with the risks spelled out first.