47 Markdown Files and the Race to Stop AI Marketing Slop

The Hype Moment: When Agents Stopped Being Chatbots

The current wave of AI attention is shifting from chatbots to agents. Claude Code, OpenAI Codex, and Cursor are no longer just autocomplete engines; they are environments where a large language model can read files, execute commands, and maintain state across a project. The problem is that raw reasoning ability does not confer domain expertise. As one SEO practitioner noted after a year of building AI agents, you cannot simply tell Claude “you are an SEO expert” and expect it to handle edge cases, workflows, or strategic nuance [2]. You have to define what that expertise means, how to think through the edge cases, and in what order to apply judgment.

This is the vacuum that coreyhaines31/marketingskills is trying to fill. Released into a landscape where marketing teams are scrambling to deploy “agentic” workflows rather than linear automations [4], the repository is a collection of markdown files—forty-seven at last count [9]—that encode marketing workflows into a format AI agents can ingest. It is pitched as a standard library for technical marketers who want their coding agents to stop hallucinating about conversion optimization and start following established playbooks.

What It Actually Is: A Standard Library, Not a Prompt Pack

The repository is built for the Agent Skills spec, an emerging convention that lets compatible agents discover and load domain-specific context. Each skill is a markdown file or folder that describes not just what to do, but when to do it, how to think about the problem, and which other skills to consult first. The repo distributes through the npx skills CLI, Claude Code’s plugin marketplace, or as a git submodule, which means it behaves like a package dependency rather than a static PDF guide [9].

The skills cover the full marketing stack: copywriting, CRO, SEO audits, programmatic SEO, A/B testing, ad creative, churn prevention, pricing strategy, and sales enablement. What separates this from a Notion page full of prompts is the structural rigor. The repo treats marketing knowledge as an interconnected system, not a bag of isolated tricks. An agent does not just receive a command to “write better copy”; it receives a framework for how copy interacts with conversion optimization, ad creative, and email sequences.

The Boring Part That Matters: A Dependency Graph for Marketing Knowledge

The most revealing detail in the README is an ASCII diagram that looks more like a software architecture chart than a marketing playbook. At the top sits product-marketing, a foundational context document that every other skill is expected to read first. Below it branch seven categories—SEO & Content, CRO, Content & Copy, Paid & Measurement, Growth & Retention, Sales & GTM, and Strategy—each containing granular skills like schema, popups, cold-email, or churn-prevention [README].

The skills cross-reference each other. copywriting talks to cro and ab-testing; revops talks to sales-enablement and cold-email; customer-research feeds into copywriting, cro, and competitors. This topology is the part that actually matters. An agent armed with these files does not just generate a headline; it checks whether the headline aligns with the product’s positioning, whether the page structure supports the CRO goal, and whether the schema markup matches the content strategy.

Without this kind of structural scaffolding, an agent tasked with writing homepage copy will default to generic best practices scraped from the internet circa 2023. It will not know that your SaaS is pivoting from SMB to enterprise, or that your highest-converting channel is LinkedIn rather than Google Ads. The product-marketing foundation is meant to prevent that drift by forcing the agent to ground itself in the company’s positioning before it generates a single line of copy. This mirrors the advice from practitioners building vertical SEO agents: start by documenting your manual workflows, then turn those workflows into structured instructions the agent can follow [2].

Built by a Practitioner in the Trenches

The project comes from Corey Haines, a marketer known for Swipe Files and the agency Conversion Factory, with a companion product called Magister that promises an autonomous AI CMO [README]. That lineage matters. Much of the AI marketing tooling on the market today—whether copy generators or ad optimizers—tends to be horizontal platforms trained on generic internet text. They produce what insiders have started calling “slop”: grammatically correct, strategically empty content [2].

Haines’s repo is essentially the opposite. It is narrow, opinionated, and practitioner-tuned. The cro skill does not vaguely suggest “improve your landing page”; it assumes you are working on signup flows, onboarding, popups, or paywalls. The ai-seo skill is not generic SEO advice; it is optimized for appearing in LLM-generated answers and AI search citations. These are the kinds of edge-case-laden workflows that only come from someone who has actually audited SaaS funnels for a living. In a field where startups increasingly expect marketers to wield technical and analytical skills alongside creative ones [12], the repo acts as a bridge between marketing expertise and agent infrastructure.

The v2.0 Mess and the Early-Adopter Tax

If the repo has a tell that it is still early, it is the upgrade path. Version 2.0 renamed seventeen skills and merged page-cro and form-cro into a single cro skill. The README warns that upgrading leaves stale folders behind—users will find both skills/page-cro/ and skills/cro/ coexisting until manually cleaned up [README]. The product-marketing context file also moved directories and changed filenames.

This friction is not a fatal flaw, but it is a reminder that the project is basically glue code. It is a well-organized layer of markdown sitting between an agent framework and a marketer’s brain. That is a genuine service, but it also means the repo’s utility rises and falls with the adoption of the underlying Agent Skills spec and the agent tools that support it. The taxonomy is still shifting, and if the foundational product-marketing file is thin or generic, every downstream skill will amplify that emptiness.

Why Vertical Skills Beat Horizontal Automation

The broader automation market is waking up to the same problem. Traditional marketing automation—Zapier, Make, n8n—excels at linear, trigger-based workflows: if a lead signs up, then send an email [7]. But as one analyst put it, the real competitive edge is no longer automating emails; it is automating intelligence [4]. Newer platforms like Lindy and Gumloop pitch “AI-native” workflows where agents reason across multiple steps, swap out LLM models for cost or quality, and act on live data [4].

Ortto, a marketing automation platform, now ships an MCP server that lets Claude or ChatGPT query customer data directly [10]. In that environment, an agent without vertical skills is just a fast intern with a big vocabulary. An agent with a structured skill library can actually interpret the data it is touching. marketingskills is an attempt to supply that interpretive layer for one specific domain. It offers a middle path: marketers can encode complex logic without writing Python, while giving the agent enough structure to reason rather than merely react.

The Outlook: Can Marketing Expertise Be Packaged?

The open question is whether marketing expertise—deeply contextual, creatively variable, and historically resistant to templatization—can truly be compressed into markdown files without losing its edge. The repo’s own v2.0 growing pains suggest the ontology is still evolving. And because the skills are essentially text files, their quality depends entirely on the depth of the thinking inside them.

Still, the bet is interesting. If agent frameworks become the default interface for technical work, then every profession will need its own standard library. Programmers have npm. Data scientists have scikit-learn. Marketers may end up with something like coreyhaines31/marketingskills: a dependency you install before asking an agent to touch your funnel. Whether it becomes the definitive standard or merely a well-curated starting point, it is one of the first serious attempts to treat marketing not as a prompt, but as infrastructure.