How to Build a Marketing Agent System With Claude Code

What We're Building

The goal is a marketing operating system driven by Claude Code agents. One primary agent acts as an orchestrator, delegating work to a team of specialists that can be operated directly through the Claude Code CLI.

By the end, the system will have:

• An orchestrator agent that handles strategy, prioritization, and cross-campaign coordination
• Seven specialist agents (research, SEO, content, email, social, analytics, campaign management)
• Five predefined workflow pipelines for common marketing operations
• Brand onboarding that captures your ICP, voice, and messaging pillars
• Persistent state management so campaigns survive between sessions
• Governance controls so each agent only accesses the tools it needs

Everything runs inside Claude Code. No external servers, no application code to maintain, no deployment pipeline. Just markdown instruction files and the Claude Code CLI.

Setup: Start From Nothing

Create a new empty folder. That's it. No boilerplate, no starter templates. Open it in VS Code with the official Anthropic Claude Code extension installed, and you're ready.

For this walkthrough, I'm using a Claude Max plan with Sonnet 4.6 in medium effort mode. You don't need Opus or extended thinking for this kind of work. Sonnet handles it well and keeps token usage reasonable.

If you need the VS Code extension, click the extensions icon, search for "Claude Code," and install the official Anthropic extension. That gives you the integrated chat interface in the editor view.

The First Prompt

Here's the prompt I used:

I would like to create a marketing operating system driven by Claude Code agents using one primary agent as an orchestrator, and leveraging a team of agent specialists that can be operated directly through the Claude Code CLI.

Simple. High-level. And that's intentional.

One important detail: I specified "Claude Code CLI" explicitly. Without that, Claude Code tends to reach for the Agent SDK, which is great for building standalone applications but isn't what we want here. We want everything operable from the command line interface directly.

Plan Mode: Architect Before You Build

Before hitting enter, I shifted Claude Code into plan mode.

Claude Code has several prompting modes that control its level of autonomy. Plan mode is read-only: it prevents Claude Code from editing any files. Instead, it creates a plan before executing anything.

What I love about plan mode is that it turns into an interview-style conversation. Claude Code asks clarifying questions, explores options with you, and builds a shared understanding of what you actually want before writing a single file. This is why starting with a high-level prompt works. Plan mode fills in the gaps through conversation.

The other modes worth understanding:

• Ask for edits — Claude asks permission before each file change
• Edit automatically — Claude makes changes but you can review
• Bypass permissions — Claude approves every decision autonomously (use with caution)

The Clarifying Questions

With plan mode active and the Superpowers plugin installed, Claude Code launched a visual companion and started asking questions:

What is the primary goal? Full stack marketing OS. Content at scale, campaign orchestration, growth and demand gen pipeline. An operating system should do everything.

Who is it for? My own company. A B2B SaaS that sells marketing agents.

Which channels? LinkedIn, email, SEO. No paid yet.

How will you interact day to day? I wanted a hybrid of high-level commands and structured workflows, plus the option for autonomous tasks. The key is orchestration: one agent delegates, routes information, and invokes sub-agents that operate autonomously.

External tools and data sources? Web research, CRM, analytics. I skipped the LinkedIn API since it's notoriously difficult to set up.

Approval model? Configurable per agent. This is critical for governance. You don't want your SEO agent sending emails to customers. You don't want your research agent creating content. Each agent should have specific tool permissions defined through prompts and hooks.

Three Architecture Approaches

After the interview, Claude Code proposed three architectural options:

A. Flat Agent Pool

One orchestrator talks directly to all specialists. Simple to build, fast to get running. But the orchestrator becomes a bottleneck, multi-step campaigns are hard to maintain, and there's no mid-tier for campaign state management.

B. Hierarchical System

Orchestrator handles strategy and cross-campaign priorities. A campaign manager sub-agent handles workflow within each campaign. Specialists execute. Clean separation of concerns, campaign state lives in the manager, scales to parallel campaigns, natural place for approval gates. More complexity upfront.

C. Workflow Pipelines

Predefined workflow templates where agents plug in as executors of each pipeline step. Highly predictable and repeatable, easy to audit. But rigid. If you have an ad hoc request, there's no predefined pipeline to accommodate it.

The Hybrid Architecture: B + C

I liked elements of both the hierarchical system and the workflow pipelines. So I asked Claude Code if there was a hybrid approach.

The answer: yes, and it's a natural fit. The hierarchy provides the intelligence layer (strategy, prioritization, routing). The workflow pipelines give the campaign manager a structured way to coordinate specialists for known operations.

The result:

• Orchestrator agent at the top, focused on strategy, prioritization, and cross-campaign coordination
• Campaign manager in the middle, managing the pipelines and maintaining campaign state
• Specialist agents at the execution layer, doing the actual work

This gives you the best of both worlds: predefined workflows for known operations (product launches, weekly content, email sequences) plus the flexibility to handle ad hoc requests through the orchestrator's natural language interface.

The Design Spec Review

With the architecture decided, Claude Code walked through five design sections for approval:

Section 1: Agent Roster

Seven agents total, each with defined roles and default approval modes. The orchestrator, campaign manager, research agent, SEO agent, content agent, email agent, and social agent.

Section 2: Workflow Pipelines

Five predefined pipelines for launch: product launch campaign, weekly content development, outbound email sequence, sprint reporting, and performance reporting. More can be added over time.

Section 3: CLI Interface

Three interaction modes: natural language (tell the orchestrator your goal), slash commands (direct commands for specific agents or pipelines), and scheduled tasks (cron-triggered pipelines).

I added a requirement here: I wanted the orchestrator to present a concise menu of options at session start, using the Ask User Question tool to kick off structured interview-style conversations. No one should have to start from a blank freeform chat.

Section 4: Project Structure

All agent definitions are stored as markdown instruction files. No application code. Claude Code invokes its own sub-agents and assigns them roles autonomously. State management through persistent files so campaigns survive between sessions.

Section 5: Brand Context and CLAUDE.md

CLAUDE.md is the master project file that every session reads on startup. It contains project rules, brand architecture (company overview, ICP personas, brand voice, messaging pillars), and a session.md file that acts as a continuous handoff document between sessions.

On first run, the orchestrator guides you through an onboarding flow to capture all of this brand information.

From Plan to Implementation

With the plan approved, Claude Code shifted from planning to implementation. I switched to bypass permissions mode so it could work without interrupting for file-by-file approval. (Only do this if you're comfortable with Claude Code making autonomous edits. Get comfortable with the permission system first.)

The result: 27 files created in one pass. Eight agents, five workflow pipelines, and eight slash commands. Everything committed and ready to run.

First Run: Brand Onboarding

To test the system, I opened a new session in the same directory. The orchestrator read the blank session.md file, recognized it was a first run, and automatically launched the onboarding flow.

The onboarding walked through seven questions:

1. Company description. What do you do, who do you serve, market position.
2. ICP details. Title, company size, pain points, language. I described solopreneurs and small agency owners dealing with execution time, resource costs, software stack bloat, and AI option overwhelm.
3. Brand voice. I went with warm and approachable.
4. Messaging pillars. I had this work done already, so I pasted in my full StoryBrand messaging framework. Claude Code called it "a gold mine" and used it to fill questions one through five completely.
5. Channel guidelines. Default for now.
6. Content guardrails. No hype or false promises. No agency bashing. A little technical jargon to establish authority and credibility.

After onboarding, the system presented a menu: what would you like to do? Active campaigns: zero. Pending reviews: zero. Ready for action.

Live Campaign Execution

I selected "plan a new campaign" and chose product launch. The orchestrator asked for a one-liner (the Builder Edition), target audience (solopreneurs and small agency teams under five), channels (LinkedIn, email, blog, SEO), and timeline (ASAP).

Then the system took over. Following the architecture exactly as designed:

1. The orchestrator established the campaign goal and dispatched the campaign manager
2. The campaign manager familiarized itself with its role and the launch campaign pipeline
3. The campaign manager invoked the research agent for a market research brief
4. Then the SEO agent for keyword research and content structure
5. Then the content agent for LinkedIn posts, a blog post, and email sequences

What the Research Agent Found

The research brief surfaced some genuinely useful insights:

Tool sprawl is the defining pain of 2026. Solopreneurs are paying for five to ten AI subscriptions and getting inconsistent, disconnected output.

The market gap is system thinking, not feature chasing. Every major AI marketing platform (Jasper, HubSpot, etc.) targets enterprise or technical teams. No credible affordable system is designed specifically for the non-technical solopreneur who wants a coherent marketing engine. I'd want to verify that claim, but it's a strong starting hypothesis.

AI agent adoption is accelerating. 56% of entrepreneurs already use AI in workflows, saving an average of six hours per week. The appetite is proven.

AI content saturation is a real anxiety. Solopreneurs are afraid of producing indistinguishable AI slop. A system that produces on-brand, consistent content is perceived as the antidote.

$147 sits in an underserved pricing tier. Comparable platforms range from $39/month to $800/month. A one-time purchase with 15 coordinated agents fills a genuine pricing vacuum.

The brief included citations for every data point. Always verify AI research, but this is a strong starting point for campaign positioning.

The SEO Brief

The SEO agent produced target keywords with intent classification and estimated volume, content structure recommendations, competitor gap analysis, and "People Also Ask" questions. Solid foundation. If you're using a dedicated SEO platform like Ahrefs or SEMrush you'll get more granular data, but web research gets you a very usable starting point.

The Content: Honest Assessment

Here's where I'll be direct: the initial content output was okay but not great.

The LinkedIn posts had the right structure and addressed real pain points, but I spotted AI-isms immediately. Em dashes everywhere. That signature three-beat sentence pattern ("No agency. No retainer. No sprawl.") that is unmistakably AI-generated. The voice wasn't wrong, but it wasn't mine either.

The blog post was better structurally. Good headers, direct pain points addressed, decent flow. A solid starting point that with better contextual input could be very good.

The email sequence needed the most work. Same em dash problem, same three-beat patterns, same generic AI cadence. To be fair, I didn't give it detailed voice and tone guidelines or example emails to model from. That's where the refinement comes in.

How to Fix This

AI content quality is a function of context quality. The system produced passable first drafts because it had basic brand information. To get to publishable quality, you need to:

• Feed it examples of your actual writing (emails you've sent, posts you've published)
• Define specific voice and tone guidelines beyond "warm and approachable"
• Set up a quality scoring system that measures readability, brand voice alignment, and SEO metrics
• Regenerate or manually refine until your quality thresholds are met

My own workflow measures every piece of content against readability, brand voice, and SEO scores. If those indicators aren't green, I regenerate or refine manually. AI is not going to get your content right without significant training, refinement, and context management.

What You End Up With

After one session, the system produced:

• A complete marketing OS with seven coordinated agents
• An orchestrator that presents structured options and guides conversations
• Brand onboarding that persists across sessions
• A full product launch campaign with research brief, SEO brief, LinkedIn posts, blog post, and email sequence
• Persistent campaign state so you can pause and resume
• Governance controls per agent

No application code. No deployment. No server. Just markdown files and the Claude Code CLI. You can open the project folder in Finder, access all the generated assets, and bring them wherever they need to go next.

Where to Go From Here

This is only the beginning. From here you could:

• Add integrations. Connect to your CRM, analytics platform, or email service through MCP servers.
• Build more agents. Add specialists for paid media, video scripting, PR outreach, or competitive monitoring.
• Improve context management. Build a local knowledge base so agents reference your past campaigns, performance data, and customer feedback.
• Refine voice and tone. Feed the system examples of your best content so it learns your actual writing style, not a generic approximation.
• Add more workflow pipelines. Define repeatable processes for quarterly planning, content repurposing, or event marketing.

The architecture supports all of this because the foundation is clean: orchestrator for strategy, campaign manager for coordination, specialists for execution.

Key Takeaways

Plan mode is your best friend. Start there. Let Claude Code interview you, explore architecture options, and create a shared plan before writing any files. The plan you get is better than whatever you would have specified upfront.

The hybrid architecture works. Hierarchical intelligence at the top for strategy and routing. Workflow pipelines at the execution layer for repeatability. You get flexibility and structure without sacrificing either.

Governance matters. Not every agent should have access to every tool. Define what each agent can and cannot do. This prevents things from going off the rails and builds trust in the system over time.

AI content needs refinement. The system will produce first drafts quickly. Those drafts need human review, quality scoring, and iterative refinement before publishing. Don't ship AI slop. Your audience can tell.

Everything is just markdown. Agents, workflows, slash commands, brand context: all markdown files. No code to maintain. Easy to version, easy to modify, easy to understand. That's the power of building inside Claude Code rather than building a standalone application.