Frequently Asked Questions About Grace Leung AI Marketing Team Builder
22 answers covering everything from basics to advanced usage.
// Basics
What is the four-layer model for building an AI marketing team?
The four-layer model is: map your marketing functions, convert each repeatable task into a skill, group skills into non-overlapping agent roles, then connect agents and skills as a coordinated team. The layers are sequential — skipping one produces unfocused, unreliable output. Each layer builds on the previous, so mapping must come before skills, and skills before agents.
What is the difference between system folders and working folders?
System folders (context, SOPs, templates) hold reusable infrastructure that pre-equips agents with brand knowledge. Working folders (ads, pages, presentations) capture what agents produce. Keeping them separate maintains clear agent orientation — mixing them breaks it. Think of system folders as the team's reference library and working folders as the output desk.
What inputs are required before I start building?
You need four required inputs: brand context files (voice guide, style guide, product offerings, strategy), branded templates for agents to study, a marketing function map of weekly repeatable tasks, and a project folder structure split into system and working folders. Optional inputs include external API keys for MCP-connected tools and a shared Notion Kanban task board.
What is a style library and how is it different from a template?
A style library is a folder of on-brand visual templates plus a style guide describing when each style fits, placed under the templates folder. Unlike an exact-copy template, it serves as visual inspiration — Claude studies the vibe and patterns rather than reproducing pixels. This produces creatively on-brand output instead of rigid, templated work.
// How To
How do I set up the initial project folder structure?
Create a root brand folder, then inside your marketing subfolder create system folders (context, SOPs, templates) and working folders (ads, pages, presentations). Think of it as how a real business organizes its office — the folder architecture is the team's physical workspace. Set this up before loading context or building anything else.
How do I install official skills in Claude Code?
Use the /plugin slash command to add Anthropic's official skills GitHub repository marketplace, then search for and install the document skills pack. Confirm installation by checking that all official skills appear under the slash command menu. These official skills become the base you extend with the Reference-Based Method to create custom branded skills.
How do I connect an external tool like an image generator to a skill?
Create a .mcp.json file in your project root declaring which MCP servers Claude can connect to and the required API keys. Restart Claude Code and verify connections via the /mcp slash command before building the dependent skill. Do this setup first so the skill has a working external connection when it needs to call the service.
How do I create a dedicated agent?
Type /agents in the Claude Code terminal, click 'Create New Agent', select 'for this project only', and let Claude generate the Agent MD file. Define the agent's name, which skills it uses, its tools, default model, color, and memory location. Each agent gets its own markdown file in an Agents folder. Keep roles non-overlapping and matched by cognitive type.
How do I write agent routing rules?
Add explicit rules to CLAUDE.md before running complex tasks, defining when Claude delegates to a sub-agent versus calls a skill directly. The rule of thumb: synthesis tasks requiring judgment (research, strategy) benefit from agents; executional, straightforward tasks (single content formats) can be handled by skills alone. These rules make the system self-directing on multi-step campaigns.
// Troubleshooting
Why is my AI team producing generic, off-brand output?
You most likely skipped context loading. Agents built without brand voice guides, style guides, and strategy documents produce generic output that misses brand standards entirely. Fix it by loading all brand context files into the dedicated context folder before building or rebuilding any skills and agents — this single step makes the biggest difference in quality.
Why is Claude losing focus on complex tasks?
You've likely piled too many skills into one agent or one conversation, the same way one person can't be writer, analyst, and designer at once. Separate skills into non-overlapping agent roles matched by cognitive type. Also check that CLAUDE.md has explicit routing rules — without them, Claude can't reliably decide when to use an agent versus a skill.
Why does my creative output look templated and rigid?
You're treating the style library as an exact-copy directive rather than visual inspiration. Reframe the style guide so Claude studies the vibe and patterns and applies them generatively, not reproducing templates pixel-for-pixel. The style library should describe when each style is appropriate, letting Claude create fresh, on-brand work instead of clones.
My remote session is degrading mid-campaign — what's wrong?
Your context window is probably full. Ignoring context limits during remote sessions degrades performance. Type 'clear conversation' to reset rather than continuing in an overloaded window. All work saves locally, so clearing won't lose output. Disconnect via the three-dot menu and archive when done, and never share the remote control link.
// Comparisons
How does this compare to hiring freelancers for each marketing function?
A Claude Code AI team runs 24/7 as an execution layer, produces a full campaign package in about 10 minutes, and keeps deliverables thematically connected — something freelancers coordinate slowly. Freelancers bring deeper judgment and originality, but for recurring, systematizable workflows the AI team delivers consistent on-brand output at a fraction of the turnaround and cost.
How does this differ from using a single Claude conversation with good prompts?
A single conversation loses focus as scope grows and starts cold each session with no persistent brand context. The AI team separates work across specialized agents, stores brand knowledge in system folders, and uses routing rules so the right agent or skill handles each task. It's the difference between one overloaded generalist and a structured, pre-equipped team.
How does the Reference-Based Method compare to blank-prompt skill creation?
Blank-prompt creation gives Claude no brand anchor, producing generic output. The Reference-Based Method provides an existing branded example, generates a detailed analysis report of its patterns, then builds the skill from that report — anchoring output to real brand standards. The result is significantly higher fidelity, with branded deck skills typically hitting 90%+ brand alignment on first run.
// Advanced
When should a task be handled by a skill alone versus a full agent?
Use a skill alone for executional, straightforward tasks like producing a single content format. Use a full agent when the task requires synthesis, judgment, or orchestration across multiple skills — such as market research or campaign strategy. Encode this distinction explicitly in CLAUDE.md routing rules so the system self-directs reliably on complex campaigns.
How many agents should an AI marketing team have?
A typical team has around five non-overlapping agents: Data Analyst, Content Creator, Market Researcher, Creative Designer, and Campaign Strategist. The number matters less than keeping roles coherent and distinct by cognitive type. Don't over-fragment or over-consolidate — each agent needs a focused, meaningful playbook of role-matched skills.
How do I connect human teammates to the AI team's workflow?
Set up a Notion Kanban board with To-Do and Complete columns and fields for task title, details, and priority. Human teammates drop tasks with priority; prompt Claude to scan pending tasks, assign the appropriate agent, execute by priority, and update status to Complete with output file paths. This board is the bridge between your humans and AI agents.
How do I keep campaign deliverables thematically connected?
Give the team a single complex, multi-deliverable brief rather than fragmented requests, and let Claude route across agents while the campaign narrative holds everything together. The quality signal is thematic coherence — research, brief, social posts, landing page, and creatives should all ladder up to one connected story. Test this on a full campaign brief before relying on it.
Should I update CLAUDE.md after the initial setup?
Yes — CLAUDE.md is a living document, not a one-off config file. Update it iteratively as you add new agents, skills, and routing rules. A common pitfall is treating it as one-time setup, which leaves routing logic stale and degrades performance on complex campaigns. Revisit it especially after creating each new agent.
Is one skill per workflow really the target?
Yes — aim for one skill per repeatable workflow. This keeps skills focused and reusable across agents, which is what makes them a reliable shared playbook. Over-stuffing a single skill with multiple workflows reduces reliability, while splitting one workflow across skills creates fragmentation. Map your weekly repeatable tasks first, then build one clean skill per task.