Four roles, one funnel

For weeks marketing was the thing that sat at the bottom of every list. I would build the product, sort the legal, clear the admin, and then it would be late and the marketing would slide to tomorrow. It is the highest-leverage work I have and it kept losing to everything else, because as one loose task competing with ten others, it always lost.

So over the past week I did the thing I should have done a month ago. I stopped treating marketing as a task and started running it as a small organization. Not more people. The same one person, me, but with the work split into clear roles, each one a seat that an agent or a tool actually sits in, and a single rule about who is allowed to talk to whom. I say “over the past week” deliberately, because I did not design this in an afternoon. It grew.

How it grew

The honest part is that I did not draw this up and then build it. It accreted. Each piece got added the week a problem made the previous shape stop working, and you can watch it happen across about seven days.

It started scattered: three product builds and no coordination, marketing being whatever happened in the chat that day.

The first real structure was a single marketing home with a freshness checker, built when my own docs kept drifting away from the live product. Then came the first campaign, where I learned to separate what marketing wants from what the build actually ships, and put myself in the middle as the relay.

Only at the end, when a five-day sprint made the coordination load real, did it harden into the operating model the rest of this post describes: a strategy seat, a tactical hub, a lead engine, the product, and me as the only wire between them.

That is the same way the rest of my system grew, the one I wrote about in how my agentic system actually works. A behavior starts loose, and structure gets added only when the looseness starts to cost something. Nothing here was clever up front. It was shaped by the stretch of it not quite working.

Four roles

The split is what makes it work. Each piece does one job and is judged on one thing.

The strategy seat sets direction. It owns the positioning, the brand voice, the canonical templates, and my own founder presence across products. It reads the results and decides where to aim. It does not run the day-to-day, on purpose, because the thing that sets the strategy should not also be in the weeds executing it. That is how strategy quietly turns into busywork.

The tactical hub runs the sprint. This is the brain of the operation. It owns the live funnel, coordinates what goes out and when, and turns the strategy into actual sends. When something about the product needs to change to move the numbers, the hub is the one that calls it.

The lead engine holds the people. Leads, the message drafted for each one, and the status of every conversation. It is the system of record for who has been reached, who replied, and who said yes. It does not freelance on strategy. It keeps the list clean and current.

The product builder ships the fixes. OneProposal is a Lovable app, so the product changes that support the campaign get made there, as prompts the build session relays. An activation fix or a tracking fix that the hub asks for lands here and goes live.

Four seats, four jobs. What keeps them from colliding is a single rule about who is allowed to talk to whom.

The human is the bus driver

Here is the rule that keeps the whole thing from turning into chaos: the spokes never talk to each other. The strategy seat and the lead engine do not coordinate directly. Everything routes through me.

And I am not just the bus that carries a message from one seat to the next. I am the driver, and I check everyone who gets on. Every handoff between agents passes through a person who reads it, weighs it against what is actually true, and decides whether it goes forward. That validation step is the control mechanism. It is the thing that stops a confident wrong assumption from boarding and riding the whole route unquestioned.

It sounds like a bottleneck, and it is the opposite. The failure mode of a multi-agent setup is not that an agent is too weak. It is that three of them start negotiating with each other, drift out of sync, and quietly compound a wrong assumption across the whole system while you are not looking. One hub, one tracker, and a single human wire means there is exactly one place where the context lives and one place where a bad call gets caught. I am not trying to remove myself from the loop. I am trying to be the only loop, so nothing happens that I did not see.

I want to be honest about why the human is still in the middle. At the current state of my infrastructure, I cannot drop a general agent into that seat and have it orchestrate and check the work the way I do. The judgment that catches a bad handoff is not something I can hand off yet. So the human wire is not nostalgia or control for its own sake. It is load-bearing, and it stays until the system is good enough to earn it away.

The other half of the rule is cadence. The hub does not report a firehose of every micro-decision. It reports at defined beats: after a timed test, at a midweek go or no-go, at the end of the sprint. Checkpoints, not a stream.

The funnel it drives

All of that exists to push one funnel: source a lead, reach them, let them taste the product for free, convert them to paid. Simple to draw, and the honest version is humbling.

When I actually read the funnel instead of admiring it, most of the traffic I had been counting turned out to be editor and preview noise, not people. The real reach is small. The truly honest baseline is zero paying customers.

Naming that out loud is what let me build a plan around the real bottleneck, which is reach, not supply. I do not have a shortage of leads or a shortage of product. I have a shortage of people who have actually heard of it. That is a distribution problem, and distribution is exactly what a funnel is for.

So the number for this week is plain: the first ten paying customers by Sunday. Not a thousand signups, not a vanity chart. Ten people who paid and did not ask for the money back. Small enough to be real, hard enough to force the machine to work.

Why a small org beats a to-do list

I wrote a while ago that you fall to the level of your systems, and this is that idea applied to the one area I kept failing at. A to-do list of marketing tasks loses to everything because it has no structure to defend it. A small org with clear seats and one rule about who talks to whom does not need willpower to keep running. The structure carries it.

If you want the general version of how this is all wired, the IDE and the agent and the layers underneath, I wrote that up separately in how my agentic system actually works. This piece is the narrow, pointed version: the same machinery, aimed at the one thing that has been beating me, and finally pointed in a direction where I can watch whether it lands.

By Sunday I will know whether the machine sells or just looks good drawn out in boxes. I will write that down too.