At some point the idea to automate it felt pretty obvious. I wanted Claude Opus specifically because quality matters here - these are real money decisions. But running it twice a day at 50-80 tool calls per session, API bill would be brutal. Using Claude Code sub through OAuth turned out to be a great hack for this. Honestly cool that Anthropic lets you use it this way.

Agent pulls data from three places: Apple Search Ads API, App Store Connect, and a lightweight SDK installed in the apps. Strategy side I took an open source ASA playbook, split it into phases, and mixed in what I learned from my own campaigns. Agent picks the right playbook depending on where app is at and keeps memory between runs.

Apple's APIs are a pain to work with and this is where vibe coding really bites you. Agent sort of figures them out but not fully - gets 90% right so you trust it. Then some edge case hits, like the reporting endpoint that puts your only report slot on a 31 day cooldown if you request it wrong, and now you're both fumbling. Either you spend a bunch of sessions on trial and error together or you go into the trenches and read the docs & code yourself. Which, let's be honest, those muscles have atrophied a bit at this point.

Still - genuinely great time to be building stuff solo. All this business grind that used to feel like a completely different job is just becoming engineering problems. Write a spec, let agent handle it, enjoy the building part. That's the dream honestly.