Build each piece in Claude Code first: a script that pulls your ad numbers into a daily sheet, the instructions the bot follows on every run, and a rules file listing what it may never touch. Then give a scheduled agent one job. Every morning it reads the sheet (not the ad account) and flags ads where frequency is rising while click rate falls, and ad sets that spent with zero leads. It can only suggest changes. Every bot reads and adds to one shared file about what you sell, who buys and what you've already tested. Any change in the ad account waits for approval, and the spending cap is set inside the ad platform, so no instruction can push spend past it. When it keeps making the same mistake, open its last 10 runs in Claude Code and fix the step that causes it.
Prompt
read this bot's instructions and its last 10 runs find the step that causes the mistake i marked propose one change and explain why wait for my approval before editing any files