We Built an Entire CMO on Claude Code: The 30+ Skills Architecture (2026)

Every "build an AI marketing team in 16 minutes" demo skips the part where it actually has to work in production. Claude Code went from developer CLI to the most-talked-about marketing stack in under a year. Search interest is up 1,005% YoY and HubSpot, Animalz, MindStudio, and Firecrawl have all published guides. We took a different path: we shipped an entire CMO platform on top of Claude Code. This is the architecture, the skills that actually move revenue, and the parts the demo videos quietly omit.

Key Takeaways

• Claude Code is a CLI tool from Anthropic that lets marketers turn repeatable workflows into composable skills, agents, and slash commands — running locally with full access to your files, APIs, and brand data.
• Search demand for "claude code for marketing" grew 1,850% quarter-over-quarter — the category is barely 6 months old and KD is still 6 (low).
• The single biggest production gap between a 16-minute demo and a real marketing system is brand-DNA injection, not model choice or prompt quality.
• A skill-based architecture (skills + subagents + plan mode) outperforms "one giant prompt" workflows because state, voice, and approval gates compose cleanly.
• We run 30+ live skills covering SEO research, article writing, social atomization, lead enrichment, AEO citation tracking, and publishing — all from one terminal.
• Claude Code does not replace your marketing team. It replaces the 70% of marketing work that is coordination, formatting, and re-typing the same brief into 5 different tools.
• The next moat in AI marketing is not a better prompt — it is a private skill library tuned to one brand's voice, ICP, and product.

What Claude Code Actually Is for Marketers (in One Paragraph)

Claude Code is a command-line agent from Anthropic that reads files, calls APIs, and runs multi-step tasks without a human watching each move. That is the definition (confirmed in Anthropic's official Claude Code documentation). Everything else is what you build on top of it. Marketers adopted it because brand context lives in your files, not a chat window. Your ICP doc, your banned-words list, your product positioning brief — Claude Code reads all of it at the start of every run. The result is repeatable output, not random output. Three primitives make this work: skills (markdown files that define a reusable workflow, like "write a LinkedIn post"), subagents (delegated specialists that each get their own context window and run in parallel), and slash commands (ergonomic triggers like /cmo:write-article that activate a skill from the terminal in one keystroke). Tie those three together and you have something closer to a composable marketing system than a chatbot. MCP servers handle the IO layer — DataForSEO, Supabase, Notion, your CMS — while skills handle the marketing logic: the brief, the voice, the rubric, the approval gate. At AI Topia, we run 30+ live skills from one terminal, covering the full pipeline from keyword research to publishing. Every one of them reads the same brand-DNA file first.

The reason this matters in 2026 is that Claude Code is no longer a developer toy — search demand for "claude code for marketing" grew 1,850% quarter-over-quarter, and the category is still early enough that most teams are running raw prompts instead of a real skill architecture. A raw prompt has no memory, no voice gates, and no cost controls. A skill has all three. That gap between a prompt and a production skill is exactly where the next section lives — and why so many "build your AI marketing team in 16 minutes" videos quietly skip it.

The 16-Minute Demo Trap (and Why It Always Breaks)

Demo videos are useful for inspiration. Grace Leung, Zubair Trabzada, and Duncan Rogoff have all built solid ones — watch their work and you will understand what Claude Code can do in a controlled environment. But controlled is the operative word. A demo optimizes for "wow in 16 minutes." Production marketing requires something entirely different: brand voice that holds across 50 articles, approval gates before anything goes live, and cost controls that stop a rogue research loop from burning $40 before you notice. None of that fits in a YouTube cut.

In 2026, the most common failure mode we see is not a bad prompt. It is operators who watched a demo, shipped something that looked great in week 1, and hit one of three walls by week 2.

The first wall is voice drift. A 5-line prompt has no memory of your brand. By article 3, the model starts writing in its own voice — hedged, generic, consultant-speak. The second wall is the missing human-in-the-loop. Demo scripts typically call a publish function directly. In production, that means AI-written copy goes live without a review step. One bad batch and you are issuing corrections to your entire subscriber list. The third wall is silent API budget overrun. A research skill that calls DataForSEO, Firecrawl, and a vector search in sequence with no cost gate will spend money invisibly. There is no alert. You find out at the end of the month.

Real production skills are 200 to 500 lines of markdown spec. That is not an exaggeration. A single /write-article skill needs voice rules, banned word lists, ICP parameters, approval gate logic, error handling for each subagent, cost ceilings per API call, and fallback behavior when a source returns empty. A 5-line prompt cannot carry that weight. The difference between a demo and a system you can trust is almost entirely in that spec.

Understanding why demos fail points directly to what the architecture needs to get right — and that is where the skill layer becomes the answer.

Our Skill Architecture: How 30+ Marketing Skills Compose

The AI Topia skill stack runs on three layers: a brand-DNA base every skill reads, pillar skills that handle the major content operations, and atomic skills plus MCP servers that handle IO. Nothing in the stack is monolithic. Skills call other skills. Subagents run in parallel. The brand DNA file is the single source of truth — change it once, every skill updates on the next run.

The bottom layer is a single markdown file: voice rules, ICP definition, banned words, content DNA excerpts, and a knowledge base index. Every skill reads this file before it does anything else. This is the production gap most demos skip. Without a centralized DNA file, voice drifts between runs, between skills, and between subagents. With it, thirty skills sound like one writer.

The middle layer is the pillar skills: /cmo:research, /cmo:brief, /cmo:write-article, /cmo:audit, /cmo:refresh, /cmo:publish-social, and /cmo:performance-seo. Each one is a 200-to-500 line markdown spec that defines inputs, outputs, approval gates, and cost controls. These are not prompts. They are contracts. The skill tells the model exactly what to produce, in what format, at what quality bar, and when to stop and ask.

MCP servers sit above the pillar skills and handle all external IO. DataForSEO (MCP) handles keyword data, SERP analysis, and rank tracking. Supabase (MCP) stores the knowledge base, brief history, and performance records. Blotato (MCP) queues and publishes social content. Notion (MCP) manages tasks and the editorial calendar. The rule is strict: MCPs are for IO, skills are for marketing logic. An MCP does not know your brand voice. A skill does.

The subagent delegation pattern is where the architecture earns its keep on long-form content. When /cmo:write-article runs, it does not write the article — it orchestrates. A researcher subagent pulls and structures the source data. An outliner subagent builds the section plan. Then five or more writer subagents run in parallel, one per H2. An editor subagent runs an AEO 10-point rubric pass on every section. An image-specialist subagent generates placement specs. An assembler subagent stitches the final document. The orchestrator never writes a word. It delegates, checks outputs, and routes failures.

Workspace isolation keeps parallel runs clean. Each article gets its own /tmp directory at job start. Research files, section drafts, edit passes, and the assembled output all live in that workspace and only that workspace. Two articles running in parallel never share a context window or a file path. This is not clever — it is the minimum viable discipline for production output at scale.

The full architecture took about four months to stabilize in 2026. The number of skills is not the point. The point is that each layer has one job, and the brand DNA file is the only place you ever touch to change the voice across all of them.

See how this architecture compares against HubSpot, MindStudio, and Cursor in the next section.

Which Claude Code Marketing Tool Is Right for Your Team?

No single tool wins every situation. The right choice depends on your team's technical comfort, your CRM stack, and whether you need agentic long-horizon workflows or guided step-by-step templates. Here is an honest breakdown of every option we have tested, including our own.

HubSpot's Marketing Against the Grain course is a strong starting point if your revenue motion already runs through HubSpot CRM. The course is gated behind a lead form, which tells you something about the intent: it is a qualified-lead funnel, not a community resource. You get structured prompts and HubSpot-native workflows. You do not get a composable skill architecture you own. If you leave HubSpot, the system does not travel with you.

Animalz and Firecrawl publish the best free tutorials for content marketers who are just starting. They will get you to a working first skill in an afternoon. The ceiling is low. Their guides treat Claude Code as a one-off productivity tool, not a composable system. Voice consistency, approval gates, and cost controls are not covered. For solo operators doing one article a week, that is fine. For a team doing 20 pieces a month, it is not.

MindStudio is the right call for teams that will never open a terminal. The visual builder is genuinely good in 2026 and it ships faster than raw Claude Code for simple workflows. The gap shows up in brand-voice depth: visual nodes make it hard to inject a full brand-DNA file the way a skill file can reference it at parse time. Cursor fits engineering-adjacent marketers who live in an IDE already. Its agentic flow support is partial and it has no native concept of a marketing skill library.

AI Topia's approach is Claude Code under the hood, pre-wired. Operators install the AI CMO platform and get the full /cmo:* skill suite on day one — brand DNA injection, subagent orchestration, MCP server connections, and cost gates included. The tradeoff is real: it is a product, not a DIY config. You give up some flexibility for a system that works in production before you've written a single prompt.

The next section breaks down the seven skills in that suite that have the highest direct impact on pipeline.

Which Marketing Skills Actually Move Revenue?

Seven skills account for 90% of the revenue-linked work we run through this system. They are not demos — each one is a production spec of 200 to 500 lines that reads brand DNA, enforces a rubric, and calls external APIs. Below is the exact map: what each skill takes in, what it calls, and what comes out.

Every skill in the suite follows the same contract: read ~/.claude/CLAUDE.md for voice rules, read brief.json for the job spec, call the right API, return a structured output the next skill can consume without human reformatting.

/research

/research runs a full SEO and AEO intake in one call. It hits DataForSEO for keyword volume and difficulty, scrapes the top 10 SERP results for entity and heading coverage, then runs a gap analysis against our AEO citation checklist. Input: a topic string and a target domain. Output: research.json — claim sources, competitor angles, KB excerpts, and a ranked keyword list. The whole run costs under $0.40 and takes under 90 seconds.

/brief

/brief converts research.json into a structured brief contract. It calls Anthropic's API with the voice rules and ICP from CLAUDE.md as a system prompt, then produces brief.json — keyword targets, section objectives, AEO patterns required, banned words, and eeat signals. No writer subagent fires until brief.json exists. That gate is what prevents voice drift from run one to run one hundred.

/write-article

/write-article is the orchestrator. It spins up one writer subagent per H2 section and runs them in parallel in isolated /tmp workspaces, then hands all section files to an editor subagent that enforces a 10-point AEO rubric — inverted pyramid, citation hooks, definitive statements, entity clarity, and six more. Input: brief.json + outline.json. Output: assembled.md ready for the publish step. A 2,700-word article clears the editor pass in under 4 minutes.

/audit

/audit pulls your live page via a Firecrawl-backed HTTP fetch, runs technical SEO checks (title tag, meta description, heading structure, schema markup), and runs the same 10-point AEO rubric the editor uses at write time. It returns a fix list with line-level diffs — not a score, not a report, a list of actions ranked by impact. Input: a URL. Output: audit.json + a human-readable fix list in markdown.

/refresh

/refresh targets rank-decaying pages. It pulls current rankings from DataForSEO, compares them against a 90-day baseline stored in Supabase, and identifies which pages dropped more than 3 positions. For each flagged page, it proposes targeted edits — new entities to add, headings to rewrite, schema to fix — with before/after diffs. You review and approve before a single character changes on the live page. In 2026, this skill alone recovered four first-page rankings in one quarter.

/publish-social

/publish-social takes one article and atomizes it into five formats: a LinkedIn post (hook + 5 bullets + CTA), an X thread (7 tweets), a Reddit post (native community framing, no brand voice), a newsletter excerpt, and a short-form video script. It queues all five to Blotato for scheduled distribution. Input: assembled.md + a target channel config. Output: five draft posts in social/ plus a Blotato job ID for each.

Does /performance-seo Actually Track AI Citation Mentions?

/performance-seo closes the loop. It pulls DataForSEO rankings, scans for AEO citations across ChatGPT, Perplexity, and Google AI Overviews, and merges Google Search Console impressions and GA4 sessions into one weekly snapshot. Output: performance.json + a plain-language brief that lists what moved, what stalled, and what to run next. The AEO citation scan catches ghost citations — brand mentions without hyperlinks — that standard rank trackers miss entirely.

See the full /cmo skill suite for the complete list of 30+ skills, including social writing, image generation, competitive scouting, and client-switch logic.

The next section covers what this stack cannot do yet — and where you will still hit a wall.

What Doesn't Work Yet in Claude Code for Marketing?

In 2026, Claude Code handles most of the marketing stack well. But four gaps will cost you time or money if you don't know them going in.

Image Generation Is Still Inconsistent

Nano Banana 2 produces solid hero images and simple social graphics. It falls apart on infographics — charts, diagrams, layered annotations. The generation is pretty but the data fidelity is wrong. We hand-design every cornerstone hero image rather than trust batch generation for critical pieces.

Long Video Costs Kill Batch Experimentation

Claude Code + Veo3 and Seedance can produce short-form video. The pipeline works. The problem is cost-per-minute. A single batch of 10 video clips runs expensive enough that you can't experiment freely. Until per-minute costs drop, use this for hero videos only — not for testing 20 angles on a campaign concept.

Real-Time Data Has No Place Here

Anything requiring sub-minute data freshness does not belong in a Claude Code skill. Live pricing, breaking news hooks, real-time inventory — these need a dedicated workflow tool with polling or webhooks. Claude Code pulls data at run time. That is fine for weekly SEO snapshots, not fine for anything where stale data causes a mistake.

Ungated Research Runs Will Hit $20+

A single /research-everything run without cost gates can blow past $20 in one session. DataForSEO calls, Firecrawl scrapes, and multi-step SERP analysis add up fast. Per-skill cost gates are not optional — they are what separates a production skill from a runaway experiment. Every skill in our suite has a hard token and API call ceiling built in.

These four limits are real, not edge cases. Knowing them upfront lets you design around them rather than discover them on a deadline. The next section shows the starter path that sidesteps all four from day one.

How to Start (Without Boiling the Ocean)

Ship one skill this week. That is the whole plan. Every marketer who stalls on Claude Code stalls because they try to replicate a 30-skill architecture on day one instead of solving one recurring task they already hate.

Day 1 — Ship One Skill

Install Claude Code and write a single skill that solves your most-repeated weekly task. Most marketers pick a content brief generator — it is tedious, templated, and immediately valuable. Your skill file is a plain English markdown spec, not code. You can have a working draft in under two hours.

Week 1 — Centralize Brand DNA

Create one markdown file: your brand DNA doc. Put your ICP definition, voice rules, banned words, and a content excerpt inside it. Point every skill to read this file at the start of every run. This one step is what separates a consistent AI output from voice drift after three articles.

Week 2 — Add a Subagent

Introduce a single subagent for parallel work. The most natural first subagent is a writer that drafts one section of a long piece while another instance handles the next. You do not need five subagents. One demonstrates the pattern and cuts long-content time in half.

Week 4 — Wire One MCP Server

Connect one MCP server for the tool your team lives in. Notion, Airtable, Google Search Console, or your CMS all have MCP connectors in 2026. Pick the one that causes the most copy-paste friction today. One connected data source makes your skills meaningfully smarter without rewriting any prompts.

Month 2 — Install a Skill Suite

Stop building from scratch. A published skill suite like cmo-lean gives you 30+ production-tested skills immediately. Install via the Claude Code plugin manager and run /cmo:research on a target topic to see the full pipeline in action. If you want to see what the full stack looks like before committing time, book a 30-minute walk-through and we will run it live on your domain.

The questions that come up fastest after this first month are exactly what the FAQ below addresses.

Frequently Asked Questions

Do I need to know how to code to use Claude Code for marketing?

No. The skill files are markdown specs in plain English. You only "code" if you want to extend an MCP server. Most marketers ship their first skill the same day they install.

What is the difference between a Claude Code skill and a subagent?

A skill is a reusable workflow (e.g., "write a meta description"). A subagent is a delegated specialist with its own context window (e.g., a writer that drafts one H2 while another writer drafts the next H2 in parallel). Skills compose; subagents parallelize.

Can Claude Code replace my marketing team?

It replaces the coordination tax — the 70% of marketing work that is re-typing briefs, formatting outputs, and chasing approvals. Strategy, taste, and customer judgment still belong to humans.

How much does running a marketing system on Claude Code cost?

Anthropic Pro/Max covers the model. Variable cost is your data layer: DataForSEO (~$0.10–0.40 per article research), Supabase (free tier holds for early stage), publishing APIs. A solo operator runs the full stack for under $200/month.

Is Claude Code better than HubSpot, Cursor, or MindStudio for marketing?

Different tools for different shapes of work. HubSpot owns CRM-native workflows, Cursor owns IDE-style edits, MindStudio owns visual no-code. Claude Code wins when you want long-horizon agentic workflows that touch your files, your data, and your APIs from one terminal.

How do you keep brand voice consistent across many AI-generated articles?

Single brand-DNA markdown file (voice rules, ICP, banned words, content_dna excerpts) that every skill reads at the start of every run. Centralized DNA is the moat — not the prompt.

Where do I download the skills shown in this article?

The cmo-lean plugin is the productized version of what we run internally. Install via the Claude Code plugin manager, then run /cmo:research or /cmo:write-article to test it on a topic of your choice.

The Takeaway

Claude Code did not become the marketing stack of 2026 because the model got smarter. It became the marketing stack because the primitives — skills, subagents, MCP servers — finally compose. A 5-line prompt cannot carry a brand. A 300-line skill that reads your DNA file every run can. The teams winning right now are not the ones with the cleverest prompts. They are the ones with a private skill library, version-controlled, that grows every week.

If you want to skip the four months we spent stabilizing this architecture, install the cmo-lean plugin and run your first /cmo:research today. Or book a 30-minute walk-through and we will show you the full stack running live on your domain.