Source: raw/Claude_Skills_-_Build_Your_First_AI_Marketing_Team_in_16_Minutes_Claude_Code.md (https://www.youtube.com/watch?v=X8afcX2s2Mo) and raw/Claude_Code_-_Build_Your_Full_AI_Marketing_Team_Agents_+_Claude_Skills.md (https://www.youtube.com/watch?v=yLXLHnD4fco), YouTube auto-captions fetched 2026-10-30.
Creator: Grace Leung (youtube.com/@graceleungyl)
Platform: YouTube
Duration: 16:43 (skills-first build) · 17:35 (agents + skills build)
Published: 2026-02-21 (skills-first build) · 2026-03-28 (agents + skills build)
Skill level: Beginner to intermediate (no coding; Claude Code Desktop in February, the VS Code extension in March)
Grace Leung, an AI-for-marketing educator, builds the same system twice, five weeks apart. In February she builds a library of five marketing skills, runs them together, hands bigger jobs to subagents and packages the library as a plugin. In March she adds agents: 12 skills grouped under 5 role-based agents, with routing rules in CLAUDE.md, a shared Notion task board and phone access through Remote Control. Both videos are sponsored by HubSpot, which supplied the free resource each one promotes. Auto-caption fixes applied: “Cloud Code” → Claude Code, “code work”/“co-work” → Cowork, “scale”/“exam” → skill/team.
Recorded February and March 2026
Several details have moved since. Cowork and chat merged into one Claude on 2026-09-16 (see Claude Cowork). Agent Teams are still flag-gated and experimental as of 2026-10-25 (see Agent Teams). Remote Control now has a server mode that serves several sessions (see Remote Control). Per-item notes are below.
Key Takeaways
- Skills hold the expertise; agents are the team members who use it. Her framing: skills contain “your best marketing knowledge, SOPs, your proven frameworks, your brand standards”. Any agent can be given any combination of skills from the library.
- Four-step design (March):
- Map your marketing function: which tasks do you do every week?
- Turn each repeatable task into a skill, “ideally one skill per workflow”.
- Group similar skills into non-overlapping agent roles.
- Connect agents and skills as a team: tell Claude in
CLAUDE.mdwho is on the team and when to use an agent versus a skill alone.
- Context first. Both builds start from a brand folder with context files (brand voice, style guide, product offerings, marketing strategy) and a
CLAUDE.mdthat tells Claude how to move through the folders. With that in place, task prompts don’t need to repeat the brand context. - Split system folders from working folders (March). System folders (
context,SOPs,templates) are reusable by every agent. Working folders (ads,pages,presentations) are where output lands. - Three ways to make a skill:
- From an SOP. Point Claude at the SOP your team already follows and ask for a skill that follows “these exact procedures”. Her research-and-strategy skill also calls Perplexity through an MCP. Her verdict on the output: “the process is actually much more important than the actual research tool that you use.”
- From examples. Give Claude your best past posts plus a storytelling framework and ask it to find the patterns. “Don’t just tell, always show your best examples.”
- From a template (March). Have Claude analyse a branded deck template and write an analysis report first. Then ask it to extend the official PowerPoint skill into a branded-deck skill. Its first 13-slide deck followed the template and was “90% done”; margins and charts still needed fixing.
- Image skills need a style library and an MCP. Her social-creative skill calls Google’s Nano Banana model through an MCP server. In March she adds a
templates/social-creatives/folder of on-brand examples, used “as visual inspiration … not the exact copy”, plus a style guide that says which style to use when. - Skills combine without being told how (February). Asked for 10 Instagram launch posts with matching creatives, Claude called the social-content and social-creative skills together. “I didn’t predefine how to split the work.” Posts and images were in a new
campaign-launchfolder in about 10 minutes. - Too many skills in one conversation is the cue for agents. “The more skills you pile into one conversation, the less focused Claude become.” Agents are “specialized team members with their own roles and tools”; skills are “the shared playbook”.
- Subagents versus Agent Teams. Subagents report only to the main Claude, which suits mostly independent tasks; she says you will use them “90% of the time.” Agent Teams share a task list, can talk to each other, and each teammate is a separate session. They suit work where agents need to review each other, but cost “much more token usage”. Ask Claude for its plan first and proceed only if you like it.
- Routing rules in
CLAUDE.mdmake the team reliable (March). After adding the agents, she updatesCLAUDE.mdwith rules for when to hand work to which agent. On a full campaign brief, Claude then used agents for research and campaign strategy (synthesis work) and skills alone for executional content. - A plugin carries the library to other brands (February). She asks Claude to package all the project’s skills, plus a custom
/landing-pagecommand that chains the research and presenter skills, into a plugin called “AI Marketing Skills”. In another brand’s project she adds it through Add plugin (from a GitHub repository or a zip upload). She also installs it in Cowork, where it followed a different brand’s colour scheme. - A shared task board lets human teammates hand work to the agents (March). Teammates add tasks to a Notion Kanban board (priority, title, details, status “To-Do”). A prompt has Claude scan pending tasks, assign an agent and execute by priority. It then moves each task to complete and records the output file paths.
The two builds
February: five skills, one plugin
Demo brand “Simple Home”, a direct-to-consumer soft-home-essentials store, run in Claude Code Desktop. She says Cowork works too.
| Skill | What it does | External tool |
|---|---|---|
| Marketing research and strategy | Follows the brand’s research SOP, writes a strategy brief | Perplexity MCP |
| Social media content | Drafts on-brand posts using extracted storytelling frameworks and patterns | Perplexity MCP (trending conversations) |
| Social creative designer | Generates visuals from post copy, with default style, palette and a prompt example | Nano Banana MCP |
| Data analysis | Takes data input, analyses it, builds a dashboard | none named |
| Campaign presenter | Turns raw materials into on-brand decks or landing pages | none named |
- Quarterly review with subagents. She asked Claude to split a quarterly review into three tasks: strategy, data analysis, and a deck combining both. Claude used the matching skills through subagents and saved a strategy brief, an interactive dashboard (click to sort and highlight metrics) and a deck in about 10 minutes.
- Agent Teams demo. Enabled in global Claude settings, then a prompt asking for “an agent team with the three specialized teammates”, where teammates one and two review each other’s work. Claude planned a team structure, a parallel phase, then cross-review and execution.
March: five agents, twelve skills
Demo brand “Go Travel”, built in VS Code so files and chat sit side by side.
- Official skills first.
/plugin→ add the official skills GitHub repository as a marketplace → install document skills. The branded-deck skill extends the official PowerPoint skill. - Project MCP config. “The most reliable way I found in VS Code” is a
.mcp.jsonfile in the project root, here holding the Nano Banana MCP server and a Gemini API key. Restart VS Code, then/mcplists the connected servers. - Agents. Created with
/agents→ “continue in terminal” → Create new agent → project only → let Claude generate the file. You describe the job, then pick skills, tools, default model, colour and agent-memory location. Claude writes a Markdown file into anagentsfolder defining the role, skills and core responsibility.- Data analyst (“thinks in numbers, charts, and patterns”). Called with an
@tag on eight campaign datasets, it produced a campaign report and an interactive, on-brand dashboard you can refresh with new data. - Content creator (“thinks in stories and headlines”), with four content skills. Asked for a campaign blog post linked to a lead magnet, it used a blog-writer skill with a keyword-research tool and a lead-magnet skill. Output: a post structured for AI search and an 11-page PDF guide.
- Market researcher, creative designer and campaign strategist, built the same way.
- Data analyst (“thinks in numbers, charts, and patterns”). Called with an
- End-to-end run. For a Japan cherry-blossom-season campaign, the team produced market research, a campaign brief (tagline “Sakura like a Local”), social posts, ad creatives and a landing page from the landing-page-builder skill, in about 10 minutes.
- Remote Control. Type
/remote-controlin the session to get a link, open it on your phone, and messages sync to the local session. “Check the task board” from the phone ran the Notion workflow. She warns not to share the link, because it gives “full control of your session”.
Try It
- List the marketing tasks you repeat weekly; circle three that already have an SOP or strong examples.
- Make a brand folder with
context/,SOPs/,templates/and an output folder, and have Claude draftCLAUDE.mdfrom it. - Build one skill from an SOP and one from examples. Test each on a real task without restating brand context.
- Run a two-skill task (posts plus creatives) and check that both outputs share one brief.
- When a conversation carries more than a handful of skills, group them into two or three agents. Add routing rules to
CLAUDE.md. - Package the library as a plugin and install it in a second brand’s folder to check what travels.
Implementation
Tool/Service: Claude Code (Desktop app in February; VS Code extension in March), Cowork for the plugin demo, Perplexity MCP, a Nano Banana MCP server, Notion.
Setup:
go-travel/
marketing/
CLAUDE.md # project instructions + agent routing rules
.mcp.json # project MCP servers (Nano Banana + Gemini API key)
context/ # brand voice, style guide, offerings, strategy
SOPs/
templates/
social-creatives/ # on-brand visual references + style guide
ads/ pages/ presentations/ # working folders
.claude/
skills/ # one folder per skill, with references/assets
agents/ # one .md per agent roleThe folder names come from her on-screen description. The .claude/ location is the standard Claude Code layout; she says only that a “skills” folder and an “agents” folder appeared.
Cost: Not stated. She notes Agent Teams use “much more” tokens than subagents. The Nano Banana MCP needs a Gemini API key, billed by Google.
Integration notes:
- Skill Creator. In February she said the official Skill Creator “is not built-in inside Claude Code” and had to be installed from Anthropic’s skills repository. The wiki’s skill-building guide describes it as shipped in
anthropics/skillsand as a Claude Code plugin; check/pluginbefore assuming it is present. - Agent Teams still need
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1(see Agent Teams). Cowork runs parallel work as subagents, not Agent Teams (see Claude Agent Hierarchy). - Remote Control is not Dispatch. She calls them “essentially the same thing”. The wiki’s comparison separates them: Remote Control drives a running Claude Code session, while Dispatch sends a task to the desktop app to work in desktop apps (see Remote Control, Dispatch).
- Her “one connected session” limit applies to
/remote-controlinside an interactive session. Server mode (claude remote-control) serves several sessions, with--spawn worktreefor parallel work. - Cowork plugin install. Since the September 2026 merge, Cowork is no longer a separate tab, so the February “bring the plugin to Cowork” step happens inside the merged Claude app.
Open Questions
- The two HubSpot resources (the Loop Marketing report and her “AI toolkit”) are sponsor material and weren’t reviewed for this article.
- She doesn’t show the routing rules she added to
CLAUDE.md, so the exact wording that made Claude pick agents versus skills isn’t recorded. - “Simple Home” and “Spark agency” are caption spellings of demo brands; neither matters to the method.
- The Notion task-board prompt is described, not shown; whether it runs through the Notion connector or an MCP server isn’t stated.
- Neither video gives token or dollar figures for a full campaign run.
Related
- Ben AI’s Eleven Marketing Skills — the nearest neighbour: a skill library for marketing with rules for content skills
- Turn Claude Into a One-Person Marketing Team (Nate Herk) — the same context-folder scaffold, with dollar costs
- Build Your First AI Agent in Claude Code Without Coding (Grace Leung) — her October single-agent build with memory and cloud routines
- A Portable AI Marketing Agent Workspace (Grace Leung) — moving this kind of setup to Codex and back
- ChatGPT Work for Marketing (Grace Leung) — her marketing agent team rebuilt in ChatGPT Work
- Claude Code Subagents — the mechanism behind her
/agentsroles - Claude Code Agent Teams — the flag-gated alternative she demos
- Plugins and Marketplaces — how her “AI Marketing Skills” plugin is packaged and installed