Let a Claude /loop
Run Your Meta Ads

The setup is almost suspiciously simple. The MCP is hosted by Meta itself – you add it to Claude or Codex as a connector, log in with your normal Meta account in the browser, and you're connected. No developer app, no API token to manage. One detail worth getting right: Meta offers three permission tiers on the consent screen – take read/write, decline the financial tier. Building campaigns never requires access to payment methods.

Here's what that looks like in practice. Running the ads for Bridesmaid, the wedding-planning startup this series builds in the open, I gave the agent one instruction: create a version-two campaign with a freshly-engaged audience, twenty euros a day, in parallel to the first one – with the current best performer as a control alongside three new ads.

Minutes later the structure existed: the campaign, the ad set with budget and conversion event, four ads – each on its own creative angle, with the tracking parameters from last week's setup baked into every link. The three new creatives? Claude wrote the angles and generated the images with an image model, in the same conversation. And critically: everything is created paused. Nothing spends a cent until a human flips it live.

Honesty about the edges: the MCP builds the skeleton – structure, budgets, tracking – and reads performance, but a few finishing touches still live in the Ads Manager UI, like picking detailed-targeting segments and some creative-polish fields. Our division of labor is simple: the agent drives the structure, and once per campaign I spend five minutes in the UI on the finish. If you want the strategy underneath the structure, the Growth Codex chapters on campaign structure and ad creatives are the frameworks this build encodes.

Setup is the convenience. This is the unfair advantage. When the same agent that runs your ads also holds your analytics stack, every performance review reads three sources at once: Meta's delivery numbers, your product analytics, and your own database – the conversion truth we wired last week. Meta isn't blind to conversions – it does get signup data back. But only for the roughly 40% of people who consent to tracking, and at a budget this small that consented sliver is far too thin to steer on. So in practice Meta leans on what it measures cleanly: the click. Your database carries no such gap – it sees every signup, consented or not, and what each one did next.

Our first campaign made the difference brutally concrete. Two ads, two stories:

Clicks are not customers. If we had optimized on Meta's surface metrics, we would have scaled the pretty ad that converts nobody. The database settled it – the same reconciling-truth pattern from the analytics setup, now making spend decisions.

And the own-data review goes one level deeper than which ad wins. Our database tracks what signups do after signing up – and it showed that most paid signups never sent a single message in the product. That's not an ad problem; that's an onboarding problem, and no amount of creative testing fixes it. Knowing which problem you actually have is exactly the discipline the Growth Codex calls what to improve – find the bottleneck before pouring more in at the top.

First, what the loop actually is. In Claude Code Desktop, /loopis a built-in command: you hand it a prompt and a cadence – /loop 12h “run my ad review”– and it re-runs that prompt on its own, waking the session on the schedule you set (give it no interval and it paces itself). The session stays alive between runs, so every wake-up still carries the context of the last. That's the whole trick: a prompt that re-fires on a schedule and remembers.

And a repeatable review is exactly the kind of prompt /loopwants. Once you've run that review interactively a few times, it has a shape – pull the numbers, reconcile the sources, judge each ad, propose the one or two changes worth making. Hand that shape to the loop and it becomes self-prompting: it wakes on your cadence, runs the review, prompts itself about what to do better, and logs everything – so every run starts smarter than the last.

The part that makes this safe to sleep on is the leash. You decide how much freedom the loop gets, on a dial: only suggest changes · update campaign settings · manage budgets within a cap · create new creatives for your approval. You place the checkpoints where decisions need you, and your guidance lives in the prompt – every correction you make becomes a rule for the next run.

Ours runs conservatively, and the rails are non-negotiable: every object the agent creates starts paused, it only ever touches campaigns it can identify by name as ours, bulk operations are forbidden, and any money change waits for an explicit yes – shown to me with the exact object and amount first. Self-improving describes the proposals, never the spending. This was born as the reporting loop in the analytics setup; now it carries the whole operation.

So what did the loop actually surface, reviewing day after day? On a few hundred euros of spend, blended cost per signup is sitting around €22 over the last ten days – four to five times our ~€5 target. The useful part isn't that number; it's that the loop traced exactly where the money leaks. Three findings, straight from the journal – early-stage and small-sample, so read them as directional signals, not settled science:

Notice what the three findings share: the top of the funnel is fine. Clicks are cheap, the creatives land, the traffic shows up. Every leak sits belowthe click – in the signup form, in the first message, in activation. Which is the whole argument for reviewing on your own data: Meta can only optimize toward the click, and the click was never the problem.

That's the shape of the week. The MCP takes the setup tax to zero – describe the campaign, get the campaign, everything paused until you say go. Your own data keeps the reviews honest – clicks are not customers, and your database is the judge. And the loop makes it relentless – on a schedule, on a journal, on a leash you set. None of it is locked to one tool: Claude or Codex, the pattern holds.

Every one of these findings is already an open workstream – and the most interesting one, that post-click drop-off, is where this series goes next. The same loop is now running A/B tests through Claude Code and PostHog to lift those signup-completion and activation rates – and what those tests are teaching us is the next Captain's Insight.

Cheers,

Ben