Building the engineering side of the Dark Factory operating model was easier than I expected, and the reason was clear in retrospect: I had 20 years of SDLC experience to draw on. I already knew what a good pull request workflow looked like. I knew what a reasonable CI/CD pipeline should do, how code review should work, what a deployment handoff needed. When I started encoding those patterns into agentic workflows, the agents fit into a framework I had already earned through practice. I was externalising existing judgement, not inventing it from scratch.

Marketing was different.

I have been involved in marketing for years. I have worked alongside good marketers, seen effective campaigns, delivered parts of it myself. But I did not have the same depth of domain expertise to design a marketing department from scratch. The patterns were not already in my head in the same way they were for engineering. I knew good from bad well enough to recognise it. I did not know it well enough to teach it.

So I took a different approach. Instead of working from instinct alone, I gathered reference material from people whose work I trusted: Alex Hormozi, Gary Vee, and others I had followed and learned from over the years. Not to imitate them. Not to pretend they work for the company. But to extract the structural logic, the content mechanics, the playbook principles that experienced operators had already developed — and then adapt those to what I am actually building.

The agents helped with extraction. I used those materials as reference inputs. The department pulled out hook patterns, campaign frameworks, offer mechanics, and proof-placement logic. That became the foundation for the playbooks, SOPs, tone guidance, and channel rules now shaping the workflow.

But extraction alone is not a department.

The first drafts from the marketing agents were technically competent and obviously wrong. The content used phrasing I would never use. Generic one-line LinkedIn spacing that signals "I've been on the internet too long." Stock AI sentence pivot patterns. The output was plausible enough to fool a casual reader but it did not sound like me. It did not sound like a product with real conviction behind it. It sounded like an agent summarising marketing advice it had read somewhere.

That is when the governance layer mattered.

I built explicit voice and tone constraints. Banned phrases. A founder-voice rewrite process that strips generic language and re-centres on real product truth and operator perspective. A review loop. A checklist. Not just "write in my voice" — but a documented, reusable set of constraints that the department applies before any content leaves the queue.

The department now has doctrine, audience definitions, campaign priorities, channel rules, content playbooks, review checklists, and handoff processes. It records what it produced, what rules it applied, what worked, and what needs to improve. That matters because the system is not just generating content. It is learning how to produce better work without drifting away from the product, the audience, or my voice.

Is it finished? No. The department is work in progress. The evidence review cycle is still maturing. The insights library is thin. There are gaps in conversion-oriented surfaces that need more structured campaign work. I am honest about that.

But it validates the broader point I have been making with the Dark Factory operating model: agentic workflows become valuable inside a company when they are wrapped in an operating model. The leverage is not that agents produce more content or more code. The leverage is that domain experts can encode judgement, standards, playbooks, and governance so agents increase output without collapsing quality.

This is not a post about replacing marketers. Existing staff become more productive when the system is designed well. Domain experts shape the work. Agents do more of the production, review, extraction, and drafting. Humans provide judgement, taste, accountability, and direction. Governance protects quality as output increases.

The same logic applies in engineering. It applies in marketing. I expect it applies in most operational domains where depth of expertise exists to encode.

The AI was the easy part. The hard part was teaching the department what good looks like.

This post was produced through the governed marketing workflow it describes. Not by a one-off prompt. By a department with documented standards, voice validation, and a review loop — running on the same operating model discipline I use for engineering.

That is the point. Not random AI content. Departments with playbooks, constraints, handoffs, and governance.