Vibe Code Your
Analytics Setup

The whole instrument is four pillars, stacked from the legal gate at the top to the permanent record at the bottom. None of them is exotic. What's new is that, because the funnel and the product now share one codebase, one agent can wire all four in a single sitting – and you end up understanding the shape well enough to trust it.

For me the four pillars are just natural. Our business home is Germany, so GDPR means a consent gate is a given, not a choice. You need one web-analytics tool – that used to be Google Analytics for me, today I prefer something more modern like PostHog. You need conversion tracking for every ad platform you run, or the algorithm is flying blind. And the fourth pillar, the database, came out of building my first SaaS, CaptAIn – the codebase that became pirateskills.com. For the first time I had a shot at extremely reliable attribution built on my own data and my own cookies, a fixed reference point I can hold Meta's numbers up against.

Which is the mental model worth keeping: the database is what actually happened, PostHog is how they behaved, and Meta is what the ad platform needs to optimise. The three will never show the same number, and that's correct – the database is the one that settles it. The Growth Codex walks the same four steps in its data analytics chapter, including the database as a tracking method in its own right.

The running example is Bridesmaid, the wedding-planning startup I'm building – and the funnel all of this measures is live, so you can walk it yourself at bridesmaid.love/tools/first-week.

I built the whole thing in one flowing session with Claude Code on Opus 4.8, in this order: consent first, then PostHog and Meta, then the database. The order barely matters – it just made sense to me to put the legal gate in before anything that fires. The agent wrote the documentation for each part as it went, so future sessions still have the full context. I never wrote a line of it by hand. I talk to the agent.

What follows is each pillar in plain language, with the prompt I used. Paste them into your own agent, in order, adapting the domain names – and put every real key and ID in environment variables, never in code.

The real point isn't that you memorise any of this. It's that you understand the shape – four pillars, a consent gate, a cookie that crosses subdomains, a de-duplicated conversion – well enough to direct an agent through it and check it did the right thing. That's the skill.

Once it's running, here's the thing nobody warns you about: the same conversion gets counted three different ways, and the three numbers never match.

Consent, ad-blockers and lost pixels mean Meta and PostHog both undercount. Your database doesn't – it's first-party and consent-independent, so it's the conversion truth you reconcile the other two against. Cost per signup is ad spend divided by database signups, sliced by campaign.

But I didn't want to read three dashboards by hand every day. So I did the obvious thing: I had the agent build the report too.

I run that report with Claude Code's /loop – a built-in command that re-runs a prompt on a set interval. I point it at one reconcile prompt and it checks in on its own every day or two: it pulls Meta, PostHog and the database, reconciles them, and hands me back one table – ad spend to impressions to clicks to trial starts – with a written read and the one or two changes worth making. It leans on the skills I've already built; it isn't a separate product, just a repeating prompt doing the job.

Look at what that one screen shows. About €118 of spend, 90 clicks, 5,665 impressions across two ad creatives. One angle is converting; the other – 21 clicks, zero signups – is a clear pause signal. And the line that proves the whole point: Meta reports two signups, my database reports four. The sources disagree, exactly as they should, and the database settles it. PostHog wasn't even reachable on that run – which the agent flagged rather than quietly dropping.

Here's the deeper reason this matters. There would have been so much I'd have needed to know upfront to brief all of this to a developer – and that's not realistic, even with a fairly deep understanding of analytics. The truth is you only catch the mistakes while real traffic is flowing. Early on I spotted that the Meta server event was arriving without its browser twin on registration, and fixed it on the spot. You run some tracking, spend for 24 hours, see what actually shows up and what doesn't, and correct – until you genuinely trust the data coming in. That state is rarely reached without both deep understanding and the ability to keep iterating. An agent gives you the second one. The Growth Codex chapters on how we report and what to improve are the framework underneath this loop.

Here's what I want you to take from this. You need to understand the setup – the four pillars, the consent gate, the cookie that crosses subdomains, the de-duplicated conversion – well enough to direct an agent and check its work. You do not need to know every line of code; tools like Claude Code and Codex handle the implementation. What you do need is to stay in regular motion: every day or two, look at whether the data is still flowing and fix whatever isn't.

That's the whole shape. Gate everything on consent. Measure behaviour compliantly. Let the ad platform optimise on a de-duplicated conversion. Keep the source of truth in your own database, then reconcile the three and act on the gap. Build the instrument before you spend, not after, so your very first euro already teaches you something.

That's the instrument. Next week I'll show you what the first hundred or two of Meta spend actually told us once it was running, and the week after, how the ad-testing setup is built on top of it. The point of this week is that none of those numbers would mean anything without the four pillars underneath them.

Your strategy is in the repo. Your story is in the repo. Your funnel is in the repo. The instrument that tells you whether any of it works belongs there too – built by the same agent, from the same context, and finally giving you the truth fast enough to act on.

Cheers,

Ben