Not in-house-I'm agency-side-but Claude Code and Codex have become most of my day now. The setup is much simpler than people make it sound. I keep a folder per client on my machine. inside that folder lives a [CLAUDE.md] file with context about the client (offer, pricing, audience, what's been shipped, what's in test), plus the relevant API tokens wired in-Google Ads, GA4, GSC, GTM, Meta, GHL. Then I just open Claude Code in the terminal and ask it stuff.
Concrete examples from this week: "pull the last 30 days of search terms for [client] and group them by intent, flag the negative candidates", "diff this client's GTM container against the snapshot from last month and tell me what changed", "read the meeting transcript from Tuesday and tell me what I promised the client by end of week", "draft four RSA descriptions for the new landing page, using the brand voice file".
you don't need to vibe-code anything for this. You just need files in a folder and API tokens. Honest setup time is an afternoon per client to wire things up the first time.
The bit that took me longest to understand is that the model is only as good as the context you give it. Same prompt against a blank Claude versus Claude with the client folder gives you very different answers. So most of the effort is in setting up that context layer once, not in writing clever prompts.
You're not as behind as you think. Most people rolling out Claude Code internally are still at the "open a chat and ask things" stage. The leap is just making the model read structured files about the work, and that's an one-afternoon problem to solve.
