Home / Guides

AI automation for integrated marketing and design content systems

The first time I tried to automate a piece of content work, it took longer than doing it by hand. I wrote a prompt, read the output, rewrote the prompt, read it again, and eventually pasted in an example of what I actually wanted. By that point I could have written the thing twice.

That's the version of AI adoption that never makes it into a case study. The tool is genuinely fast and free of grammatical errors. What it costs you is the time spent tweaking the output into something you'd put your name on: in your voice, against a real business goal, and with some life left in it.

What fixed it for me wasn't a better model. It was building a system around the model.

What does "AI in marketing" actually mean?

Everyone in integrated marketing says "AI" now, but ask 5 people what they mean and you'll get 5 different answers. Some say efficiency. Some say better ideas. Some are just following what everyone else is doing without a clear reason.

It's true that AI can automate a lot in a second and write with technically perfect language. So why does it so often make things slower? Without guidance, that speed cuts both ways. First, it can be more inefficient, because every single task needs its own prompt written from scratch, so every time the output can be different. Second, the output drifts from the original business purpose, with quality swinging wildly between one piece and the next. That's why treating AI as a chatbot is such an inefficient way to use it.

That second problem now has a name. Merriam-Webster named "slop" its 2025 Word of the Year. They defined it as low-quality digital content produced at scale by AI. The American Dialect Society voted it their Word of the Year too, in early 2026. Two separate language authorities landed on the same word for the same problem: marketing materials produced quickly with AI and jarring to the human eye.

The scale of the gap is real. 78% of multinationals now deploy AI-generated or AI-enhanced creative across channels, according to World Federation of Advertisers research. And 96% of marketers have seen content demand at least double in two years, in an Adobe survey of more than 1,600. Brands can now generate content faster than they can see, understand or control it, which is exactly the gap governance has to close.

To make AI automation work, you need a system

The fix for AI slop isn't a smarter model, as unclear prompts only produce unreliable output. What we need is a system that governs what gets automated and which guides that automation follows, bound by a human oversight framework, or as it is usually called, human in the loop. A guardrail well built once will save you from daily rewriting.

Start with one overarching instruction file. Think of it as your project's CLAUDE.md. Keep it specific but not exhaustively long. If a particular part needs real depth, that depth belongs in a separate guide it points to, so this main file doesn't get bloated and slow down your AI model.

Then build the specific guides that govern how things get written. Your voice guide matters most here. It's your formalised, custom editorial style guide: what you say and how, your tone, and where relevant how you sound different in different contexts. This goes hand-in-hand with an anti-AI-writing guide and a design system guide. The former specifies which language patterns to avoid (like the now-notorious "not X, but Y"); the latter governs how your writing sits inside the wider UI layout. Between them, these guides draw a hard line between acceptable output and unacceptable output.

Next come skills. Most projects span genuinely different disciplines: lead generation, SEO, design, UX, social media and paid ads. Each one needs its own skill file in a dedicated directory. That's where the AI reads what needs doing in that discipline and how to do it. It only loads when the task calls for it.

One practical tip: you don't need to be an expert in every one of these fields yourself. There are plenty of resources online for building these skill files, so you can borrow the expertise instead of holding all of it in your head at once.

Separate orchestration from execution

Anthropic's own engineering team frames this underlying discipline well. It's "the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference, including all the other information that may land there outside of the prompts." That's the difference between writing a prompt and building a system.

Beyond the guides and skills, keep planning and doing in two separate passes. Orchestration is a set of instructions or documents describing what needs to be done and how. Execution is when that gets done. If that sounds like servant leadership applied to tooling, that is more or less what it is, and the Agile guide covers the human version. Anthropic documents a closely related pattern in its own engineering work: an orchestrator-workers workflow. A central process breaks the task down and delegates the pieces, rather than asking one pass to plan and build at the same time. Keeping the two separated is what makes the output more specific and more precise. A planning pass that also has to produce finished work tends to do both jobs worse.

Without data, you're flying blind

None of this matters much if your system has no real data feeding it. Connect Google Search Console, Google Analytics, Google Ads, and a tool like Ahrefs or SEMrush. You do it through MCP, the Model Context Protocol Anthropic introduced in November 2024 as an open standard for wiring AI systems to external data. That turns your system from something guessing at what works into something that can see it.

How this page itself got built

I'll confess something: I'm not a website designer, and I'm not (yet) a product manager. I built my own content workflow and this guide using exactly this system. It's what lets me create this guide, the website you're reading, and even my LinkedIn posts.

None of it was new to me. I'd built a version of this before, at scale. I moved a 16-person team off ad-hoc AI use onto a structured system, with scenario-specific prompts and a mandatory human review layer. It lifted content production velocity by more than 70%, and the system was adopted across close to 40 domains and by marketing leadership. The governance layer is what made that number possible.

For this page, I started with my voice guide and my orchestration layer: what I needed done and how, which design system I was using in Figma, and where I still needed more guides. That included how not to end up with AI slop, which mostly comes down to deciding upfront where human oversight is non-negotiable. Every task got a role: designer, writer, SEO. Then I fed in enough context to work with, from SEO requirements to the actual anti-AI instructions to outlines for every page on the sitemap.

Next, I created the Figma design system: fonts, layout, aesthetics. You'd be surprised how easy this part is. Give it a few references and some UI-element prompts you can find online and it comes together quickly. After that came execution. With the instructions and guardrails ready, and the Figma MCP installed, I could get the AI to build every page, then the whole site wireframe. I'd tweak it, send it back, and eventually commit the result and deploy it.

For storing everything and keeping it ready to go, my go-to is GitHub. You can share the repository with others. You can keep versions and roll back changes. And it's far easier for a CMS or a developer to draw from a committed system than from a folder of loose files. After that, I store finished work in Notion, which doubles as a reference the AI can draw from later.

{
  "id": "build-system-hierarchy",
  "type": "flowchart",
  "anchor_heading": "How this page itself got built",
  "title": "The build system, top to bottom",
  "orientation": "vertical",
  "layers": [
    {
      "level": 1,
      "label": "Governing instructions",
      "nodes": [
        { "id": "claude-md", "label": "CLAUDE.md", "note": "Overarching context, read every session" }
      ]
    },
    {
      "level": 2,
      "label": "Guides",
      "nodes": [
        { "id": "voice", "label": "voice.md", "note": "Editorial style guide, formalised" },
        { "id": "anti-ai", "label": "anti-ai.md", "note": "Banned words, constructions, tells" },
        { "id": "seo-geo", "label": "seo-geo.md", "note": "Headings, schema, meta, crawlability" },
        { "id": "design-system", "label": "design-system", "note": "Figma tokens, type scale, components" }
      ]
    },
    {
      "level": 3,
      "label": "Skills",
      "nodes": [
        { "id": "skill-seo", "label": "SEO" },
        { "id": "skill-ux", "label": "UX writing" },
        { "id": "skill-social", "label": "Social" },
        { "id": "skill-design", "label": "Design" },
        { "id": "skill-leadgen", "label": "Lead gen" }
      ]
    },
    {
      "level": 4,
      "label": "Working split",
      "nodes": [
        { "id": "orchestration", "label": "Orchestration", "note": "What needs doing, and how" },
        { "id": "execution", "label": "Execution", "note": "Doing it, against the plan" }
      ]
    },
    {
      "level": 5,
      "label": "Data in",
      "nodes": [
        { "id": "gsc", "label": "Search Console", "via": "MCP" },
        { "id": "ga", "label": "Analytics", "via": "MCP" },
        { "id": "ads", "label": "Google Ads", "via": "MCP" },
        { "id": "ahrefs", "label": "Ahrefs / SEMrush", "via": "MCP" },
        { "id": "figma", "label": "Figma", "via": "MCP" }
      ]
    },
    {
      "level": 6,
      "label": "Storage and handoff",
      "nodes": [
        { "id": "github", "label": "GitHub", "note": "Version control, sharing, CMS/dev handoff" },
        { "id": "notion", "label": "Notion", "note": "Finished work, reusable as AI reference" }
      ]
    }
  ],
  "edges": [
    { "from": "claude-md", "to": "voice", "type": "points-to" },
    { "from": "claude-md", "to": "anti-ai", "type": "points-to" },
    { "from": "claude-md", "to": "seo-geo", "type": "points-to" },
    { "from": "claude-md", "to": "design-system", "type": "points-to" },
    { "from": "voice", "to": "orchestration", "type": "feeds" },
    { "from": "anti-ai", "to": "orchestration", "type": "feeds" },
    { "from": "seo-geo", "to": "orchestration", "type": "feeds" },
    { "from": "design-system", "to": "orchestration", "type": "feeds" },
    { "from": "orchestration", "to": "execution", "type": "hands-off-to" },
    { "from": "execution", "to": "github", "type": "commits-to" },
    { "from": "github", "to": "notion", "type": "archives-to" },
    { "from": "gsc", "to": "orchestration", "type": "informs" },
    { "from": "ga", "to": "orchestration", "type": "informs" },
    { "from": "figma", "to": "execution", "type": "informs" }
  ],
  "caption": "Instructions flow down, data flows in sideways, and finished work exits through GitHub into Notion."
}

The thing I'd push back on hardest is the idea that any of this requires being technical. I'm a content person. I built the system described above because writing every prompt from scratch, every time, was slower than doing the work myself, and because the output kept drifting away from what the business actually needed.

What fixed it wasn't a better model or a clever prompt. It was writing down what I already knew: how the brand sounds, which words never make it past me, what a page needs structurally, where a human has to look before anything ships. Once that existed as files rather than as instructions I retyped daily, the AI stopped guessing and started working against a spec.

That's the whole trick. The system encodes your judgment so the tool can apply it consistently at a volume you couldn't hit alone. Skip that step and you get output faster, which sounds good right up until you're reviewing it.

The system is the leverage

Better models will keep arriving regardless of what any single team does. The actual advantage available right now is the governance layer around the model: the guides, the skills, the orchestration and execution split, and the real data feeding all of it. That's the part a bigger model doesn't replace.

Sources

• Word of the Year 2025: Slop, Merriam-Webster
• 2025 Word of the Year Is "Slop", American Dialect Society
• Brands struggle with AI disclosure as usage surges across marketing, Marketing-Interactive (reporting World Federation of Advertisers research)
• Adobe research finds demand for content will grow 5x by 2027, Adobe
• Building Effective AI Agents, Anthropic
• Introducing the Model Context Protocol, Anthropic
• Using CLAUDE.md files, Claude by Anthropic