There is a specific moment when you realize AI has stopped being a novelty in your marketing workflow. It happens the first time you stop running prompts one at a time and instead watch a small team of specialized agents pick up tasks, delegate, and hand work back to you already 90% done. Grace Leung recently walked through exactly that setup in Claude Code — five agents, twelve skills, a shared Notion task board, and a connection to her phone so the team keeps working while she is away from her desk. I watched the whole thing with one question on my mind: if you actually build this AI marketing team, where do the posts go?
Because an AI marketing team that can research, write, design, and analyze is only useful once someone (or something) actually publishes the work. That gap is where tools like Postiz come in, but we will get to that. First, let us look at the architecture Grace uses, because it is the cleanest framing I have seen for turning Claude Code into a real marketing operation.
Manage all your social media in one place with Postiz
Most marketers trying to work with AI agents do it backwards. They pick an impressive agent framework, bolt on a couple of MCP servers, and then look for marketing tasks to feed it. Grace’s approach flips that. You start with your actual marketing function, then build skills and agents around it.
The four steps are simple enough to write on a sticky note:
Map your marketing function. List the tasks you genuinely do every week — blog posts, decks, ad creatives, campaign reports, social carousels, newsletters.
Turn each repeatable task into a skill. One skill per workflow. Skills are reusable playbooks Claude can call when it encounters that kind of task.
Group similar skills into non-overlapping agent roles. A data analyst agent, a content creator agent, a market researcher, a creative designer, a campaign strategist. Dedicated agents produce sharper output than one generalist.
Connect agents and skills as a team. Set up your project so Claude Code knows who is on the team, which skills each agent owns, and when to reach for a skill directly versus delegating to an agent.
Grace built hers for a fictional travel brand called Go Travel, but the pattern is the same whether you are running a SaaS, a creator business, or a content agency. Inside her project folder she separates system folders (context, SOPs, templates — reusable by every agent) from working folders (ads, pages, presentations, reports — where outputs land). Loading the system folders with a brand voice guide, a style guide, product offerings, and a marketing strategy is the step that pays off every single day afterwards. Agents without brand context are generic. Agents with it sound like your team.
Skills are where you start, not where you stop
The first skill Grace builds is a branded deck skill. She uses what she calls a reference-based method: drop a clean template into a templates folder, have Claude analyze it and generate a structural report, then extend the official Anthropic PowerPoint skill with that analysis. Within a few minutes Claude is producing thirteen-slide campaign decks that follow the exact brand template — not a generic “corporate” aesthetic, but her slides.
The second skill is more interesting because it crosses into external tools. She wires up the Nano Banana MCP server via a .mcp.json file in the project root, loads a small library of on-brand social creatives as visual references, and then builds a social creative designer skill on top. Now Claude can generate a five-slide Instagram carousel with the right vibe — not a copy of any single template, but something that clearly belongs to the brand.
This is the pattern for almost every marketing skill worth building: a small reference library, a clear description of what “good” looks like, and a connection to whatever specialized model or service does the heavy lifting. Repeat that twelve times and you have a skill library that covers most of your weekly workload.
When a skill is not enough, an agent takes over
Here is the tension: the more skills you pile into a single conversation, the worse Claude gets at choosing between them. Grace puts it clearly — “just like one person trying to be your writer, analyst, and designer all at once.” Context dilution is a real problem.
This is where sub-agents come in. Agents are specialized team members with their own role description and their own curated set of skills. Skills are the shared playbooks any agent can use. Creating a new agent in Claude Code is a two-minute exercise: type /agents, create a new agent for the project, describe what it does, pick the skills it owns, pick a model, and Claude writes the agent’s markdown file for you.
Grace ends up with five agents:
Data Analyst — pulls together campaign reports and interactive performance dashboards from raw data sets.
Content Creator — owns the blog writer skill, lead magnet skill, and other long-form skills.
Creative Designer — owns the social creative, deck, and landing page skills.
Campaign Strategist — coordinates everything into a cohesive brief.
The trick that makes this actually reliable is updating the project’s CLAUDE.md file with explicit routing rules. Claude Code reads that file on every run, so telling it “use the data analyst agent whenever the task involves synthesis of campaign metrics” is what turns a folder of agent definitions into an orchestrated team.
The first time the team builds something together
The moment that sells the whole idea is when Grace gives the team a single complex task — a Japan cherry blossom campaign requiring market research, a campaign brief, social posts, ad creatives, and a landing page — and lets Claude decide which steps need agents versus which can run on skills alone. Research and strategy go to agents. Pure execution tasks stay as skill calls. Ten minutes later she has a complete campaign package, all tied back to the same theme, all on-brand, with the landing page rendered and ready.
That is the first “this actually works” moment. The second one comes when she stops being the input source entirely.
Making it real: a shared task board
An AI marketing team that only works when you are typing prompts is still a bottleneck. Grace’s solution is a shared Notion board where anyone on the real team — a human teammate, a contractor, herself from her phone — can drop tasks with a title, priority, and status. A single prompt tells Claude to scan the pending tasks, assign the right agent, execute by priority, and update the status when finished.
This is the part that turns a clever setup into a collaboration layer. A content lead can drop in “write a launch post for our April release” on Monday morning, and by the time she is back from her first meeting the draft, the carousel, and the research notes are all sitting in the project folder with the task marked complete.
Taking the team with you
The last piece is Claude Code’s remote control — you activate it on your desktop session with /remote-control, open the link on your phone, and now the Claude mobile app is talking directly to your local session. “Check the task board” from a coffee shop triggers the same flow that would run from your desk. You can send new tasks, review outputs, and nudge agents while you are away. It is the closest thing I have seen to a genuine 24/7 marketing team without hiring a 24/7 marketing team.
The missing piece: where the posts actually go
Here is what I noticed watching the full walkthrough. The team can research audiences, write blog posts, design carousels, generate ad creatives, and draft landing pages. What it does not do — at least not out of the box — is publish anything. The outputs sit in the project folder as markdown files, PNGs, and HTML drafts. Getting them scheduled into Instagram, LinkedIn, X, TikTok, YouTube, Threads, Facebook, Reddit, Pinterest, and twenty other channels is still a human job.
That is the publishing layer we built Postiz for, and it slots into a Claude Code setup like this cleanly. You have three ways to connect.
1. The Postiz MCP server
If your agents already talk to MCP servers (Grace wires up Nano Banana the same way), adding Postiz is a drop-in. The Postiz MCP exposes tools any Claude agent can call:
integrationList — lists every connected channel with its ID, platform, and display name.
integrationSchema — returns per-platform rules: character limits, required settings, available data-fetching methods (Reddit flairs, YouTube playlists, LinkedIn companies).
integrationSchedulePostTool — schedules posts across one or many channels. Supports schedule, draft, or now, handles thread-style comments, enforces per-platform character limits, and validates HTML.
triggerTool — calls platform-specific methods dynamically (fetch the subreddit’s flairs before posting, pull a list of YouTube playlists, etc.).
generateImageTool and generateVideoTool — generate assets and upload them to the Postiz CDN in one step.
In the AI marketing team setup, this means your content creator agent does not need a separate “publish” skill. The MCP tools are available to the agent directly, and it can schedule a finished draft the moment it is written.
2. The Postiz CLI
If you prefer a skill that shells out to a command-line tool — which fits neatly into the reference-based method Grace uses — the Postiz CLI covers the same surface area:
Everything returns JSON, so you can chain it into an agent that uploads a generated carousel, schedules it across Instagram and LinkedIn at a specific time, and then writes the post IDs back into the Notion task board.
3. The Postiz Public API
If you want a campaign strategist agent to plan a week of posts across twenty-eight channels in a single call, the Public API is the cleanest route. POST /posts accepts an array of integration IDs, content, media URLs, and per-platform settings. POST /upload returns a CDN URL from uploads.postiz.com — important, because platforms like TikTok, Instagram, and YouTube reject external media domains during publishing. Analytics endpoints then feed clean data back to the data analyst agent the next morning for its weekly report.
The pattern most teams land on is: the content creator agent drafts, the designer agent produces the visual, Postiz handles the schedule, and the data analyst agent reads the results back through the analytics endpoints. Three agents, one publishing layer, no human clicking “post” at 8am.
What to build first
If you are copying Grace’s architecture, do not try to build all twelve skills and five agents in one weekend. The version of this that works is boring: pick the two tasks you do most often — for most marketers that is “write a blog post” and “design a carousel” — build a skill for each, wire a single agent around them, and add a publishing step through the Postiz MCP or CLI. That alone replaces hours of every week. Then grow the team from there: a data analyst agent once you have campaign data worth analyzing, a market researcher when you are launching into a new segment, a strategist when the calendar gets crowded.
The magic is not in the number of agents. It is in the separation of concerns. A focused content creator agent with a sharp brand voice beats a five-agent mega-team with muddled context, every time.
Give your AI marketing team a publishing layer
You can absolutely build the Claude Code team from the video without ever touching Postiz — the framework stands on its own. But if you follow the walkthrough, you will hit the same wall I did: great drafts piling up in a folder, and no clean way to get them scheduled across every channel your brand lives on. That is exactly the problem Postiz solves, and the three integration paths above mean you can plug it into whichever version of the team you are building — MCP for agents that already speak MCP, CLI for skills that shell out, public API for the agents that want direct control.
If you want to see how it fits with an AI agent workflow before committing to anything, try Postiz for free — connect your channels, let your Claude Code agents publish into real accounts, and close the last gap between an AI marketing team that drafts and one that actually ships.
Discover 9 proven content marketing best practices for 2026. Learn how to plan, optimize, distribute, and measure content that drives traffic, engagement, and ROI.
