Latest / AI Fire Daily / #12 Robin: From Chatbot to Agency - How to Build a Multi-Agent Marketing Team Directly Inside Claude Code
Transcript
- 0:00We're witnessing a profound paradigm shift right
- 0:02now. We're moving away from treating AI as a
- 0:06simple chatbot. We're architecting autonomous
- 0:09five -person marketing agencies. Yeah. And we're
- 0:12doing it inside a single folder. Beat. Welcome
- 0:17to the Deep Dive. Today, we're exploring a radically
- 0:21different way to work. We're completely rethinking
- 0:24human -computer interaction here. I mean, it's
- 0:27not about writing one perfect prompt anymore.
- 0:29It's about building a structured digital ecosystem.
- 0:32If you're listening to this, you've probably
- 0:34spent hours wrangling a generic AI. Oh, yeah.
- 0:38You try to get it to write just one decent email,
- 0:40and it usually ends up sounding... Right, and
- 0:42then it starts with, you know, in today's fast
- 0:44-paced digital landscape. It drives people insane.
- 0:46It really does. Our mission today is highly specific.
- 0:49We're going to unpack exactly how to turn Claude
- 0:51Code into a repeatable marketing system. We'll
- 0:54use a test case of a travel brand. The brand
- 0:56is called Go Travel. We'll look at how this system
- 0:58builds an entire campaign autonomously. Beat.
- 1:02Okay, let's unpack this. I am incredibly ready
- 1:05for this. The framework we are looking at is...
- 1:08Brilliant. It completely solves the blank page
- 1:11problem. Before we can build the software environment,
- 1:13we have to conceptually define our employees.
- 1:15Right. We're building an AI agency. So we need
- 1:18to understand the fundamental difference between
- 1:20a skill and an agent. This is the core rule of
- 1:23the entire setup. You know, people try to automate
- 1:25their entire business on day one. Yeah. They
- 1:27end up with an AI that's like. aggressively replying
- 1:30to their mom's emails, you have to start with
- 1:33the foundational element. If we're setting up
- 1:34this system, we start with skills. Exactly. A
- 1:37skill is a repeatable workflow. It's a highly
- 1:39specific playbook. You don't just ask the AI
- 1:42to write a blog post. You create a custom blog
- 1:45writing skill. Walk me through what that actually
- 1:48looks like mechanically. It's a document containing
- 1:51specific SEO rules. It includes your exact brand
- 1:55voice guidelines. It has internal linking rules.
- 1:58It even has an editing checklist. Wow. The AI
- 2:01doesn't guess anymore. It reads that file and
- 2:04follows your predefined process. It's like a
- 2:06restaurant. The skills are the exact recipe cards.
- 2:09Yeah. And the agents are the specific chefs working
- 2:12the stations. That's a perfect way to visualize
- 2:14it. The agents are the specialized team members.
- 2:17They execute those specific recipes. You might
- 2:19have a market researcher. You might have a creative
- 2:21designer. They have entirely different cognitive
- 2:24jobs. Hovering above all of this is something
- 2:26called a clay .md file. Right. A .md or Markdown
- 2:31file is a basic text document that strips away
- 2:35messy formatting. Okay. This specific file acts
- 2:38as the general manager. It gives the overarching
- 2:41instructions. It tells the chefs which recipes
- 2:43to use. I have to push back here. Sure. Why do
- 2:46we need separate agents? Yeah. Why not just ask
- 2:49one super smart AI to do everything at once?
- 2:52Because mixing complex tasks leads to massive
- 2:55logic degradation. If one chat window tries to
- 2:58write content, analyze data, and build strategy
- 3:00simultaneously, it loses focus. Its attention
- 3:04mechanism gets scattered. I see. An agent must
- 3:06have one clear, isolated responsibility. So the
- 3:10market researcher only finds trends. They absolutely
- 3:13do not design the ad visuals. Right. And just
- 3:15like a real kitchen, if the fry cook tries to
- 3:18bake the souffle, it's a disaster. Yeah. The
- 3:21creative designer doesn't touch the campaign
- 3:23data. You keep the roles strictly separated.
- 3:27How do you actually focus the AI so it doesn't
- 3:29wander off task? Well, you define strict limitations
- 3:32clearly in its setup file. You tell it exactly
- 3:35what it should never do. So you're essentially
- 3:37building a highly controlled digital assembly
- 3:39line. You are. And it works beautifully. Now
- 3:43that we know our team's roles, where do these
- 3:45digital chefs actually work? Right. They need
- 3:48a highly organized kitchen. They need a permanent
- 3:50workspace. For this setup, the source heavily
- 3:53recommends using VS Code. That's Visual Studio
- 3:56Code. A popular app where developers write code
- 3:58and organize files. Exactly. Beat. But why use
- 4:01a coding app instead of a normal browser chat
- 4:03window? Because a chat window is basically a
- 4:06goldfish. It forgets everything. Then this code
- 4:08gives the AI spatial awareness. Cloud code can
- 4:11actually read your local folders. It can see
- 4:14your files, your skills, and your final outputs
- 4:17all in one place. If I understand the source
- 4:20correctly, we're dividing this digital workspace
- 4:23into two distinct folder types. System folders
- 4:27and working folder. That's the architectural
- 4:29secret. Think of system folders as the AI's long
- 4:32-term memory. What actually goes into those system
- 4:35folders to build that memory? This is where you
- 4:38put a context folder, an SOP folder, and a templates
- 4:41folder. This teaches the AI about your business.
- 4:44You drop in your brand voice guide. You add your
- 4:47ideal customer personas. You include your current
- 4:50marketing strategy. So GoTravel, our test brand.
- 4:53needs the AI to deeply understand its target
- 4:55demographic. Right. It needs to know the exact
- 4:58tone to use for adventurous millennials. Without
- 5:00that context folder, the AI just defaults to
- 5:03generic corporate speak. With it, the AI sounds
- 5:07like your best copywriter. I have to offer a
- 5:09vulnerable admission here. Oh. I still wrestle
- 5:12with prompt drift myself when staring at a blank
- 5:14chat window. It's incredibly frustrating. It
- 5:17happens to everyone. I mean, prompt drift is
- 5:19when the AI forgets original instructions over
- 5:22a long chat. By the 10th prompt, it's completely
- 5:25lost the plot. But if we're using the system
- 5:27folder structure, the context becomes permanent.
- 5:30Exactly. Because VS Code reads that claudeu .md
- 5:34file automatically in the background. It acts
- 5:37as a persistent system prompt. The AI physically
- 5:41cannot forget your brand. Those rules are injected
- 5:44into its memory every single time you hit enter.
- 5:47So if the system folders act as the AI's long
- 5:50-term memory, where does it put its actual daily
- 5:53work without cluttering up its brain? That's
- 5:56where the working folders come in. You create
- 5:57folders named ads, pages, reports, and social
- 6:00creatives. Okay. This is where Claude saves the
- 6:03polished outputs. And the ultimate routing manager
- 6:05for all these folders remains that manager file.
- 6:08Yeah. It tells the AI exactly what the project
- 6:10is about. It maps out how the folders are organized.
- 6:12Right. It tells the AI exactly where new files
- 6:14should be saved. How do you prevent context overload
- 6:16when the AI scans all these different folders?
- 6:20You configure the manager file to direct the
- 6:22AI only to specific context files when performing
- 6:25specific tasks. So the routing file acts as a
- 6:28protective cognitive filter. That's a great way
- 6:30to put it. The digital workspace is built. Before
- 6:32hiring the specific agents, we need to stock
- 6:34the shelves with our workflows. Yes. We need
- 6:37to craft the actual playbooks. The skills library.
- 6:40This is my favorite part. This is where you store
- 6:43your repeatable workflows so you never have to
- 6:45explain a task twice. The source strongly advises
- 6:49starting small here. Start with just three to
- 6:52five tasks you repeat weekly. People always try
- 6:54to build 50 skills on day one. It's a classic
- 6:57mistake. Build one skill, test it, refine it,
- 7:01make sure it works perfectly before moving on.
- 7:03Let's look at a specific playbook from the source
- 7:05material, the branded deck skill. This completely
- 7:08changes the game for creating presentations.
- 7:10Instead of getting a generic slide layout, you
- 7:13teach the AI your exact corporate style. How
- 7:16does an AI actually study a corporate style?
- 7:19Well, you drop a finished PowerPoint template
- 7:21into the templates folder. You ask Claude to
- 7:25analyze its underlying structure. Okay. It reads
- 7:28the XML code. It analyzes the specific hex colors.
- 7:32It studies the typography and spacing rules.
- 7:35It essentially creates a reverse -engineered,
- 7:37detailed report on what makes your deck feel
- 7:39uniquely on -brand. Exactly. Then you save that
- 7:43report as a permanent skill. The next time you
- 7:46need a strategy deck, the AI automatically applies
- 7:49those exact design parameters. That is a massive
- 7:51time saver. Huge. There's also the social creative
- 7:54design skill. This one functions slightly differently.
- 7:57Yeah, this one is about capturing a vibe. You
- 8:00store strong visual references in an examples
- 8:02holder. Then you pair it with a style guide document.
- 8:05It studies the image direction and the visual
- 8:07mood. It doesn't just copy a single template
- 8:09rigidly. Right. But doesn't forcing the AI to
- 8:13study past examples. kill its creativity and
- 8:16make the output robotic. What's fascinating here
- 8:18is that constraints actually breed creativity.
- 8:21Yeah. When you give the AI a defined sandbox
- 8:25of approved colors and moods, it stops generating
- 8:28generic stock photo aesthetics. It starts combining
- 8:32your specific brand elements in highly creative,
- 8:35unexpected ways. Freedom within a strict framework.
- 8:38Exactly. That makes a lot of sense. The AI needs
- 8:40boundaries to be truly clever. How do you handle
- 8:43updating these skills over time when your brand
- 8:47strategy changes? You simply open the specific
- 8:50skill text file, rewrite the outdated step, and
- 8:53the AI instantly adopts a new process. So dynamic
- 8:56skill files prevent the system from getting stale.
- 8:58They make evolving your business incredibly seamless.
- 9:01Our playbooks are officially written. Now we
- 9:03use the slash agents command to hire the specialized
- 9:06talent to read them. We're finally assembling
- 9:07the AI team. For GoTravel, the test system created
- 9:10five highly specific agents. First up is the
- 9:13market researcher. This agent dives into audience
- 9:16insights, seasonal friends, and competitor angles.
- 9:19Second is the campaign strategist. Right. They
- 9:21take that raw research and turn it into actionable
- 9:23campaign ideas. offers and messaging. Third is
- 9:26the content creator. They write the actual deliverables,
- 9:30blog posts, social captions, and lead magnets.
- 9:32Fourth is the creative designer. They handle
- 9:35the visual direction and specific ad creatives.
- 9:38And finally, the data analyst. Yeah. They read
- 9:41the post campaign data and build performance
- 9:43dashboards to see what worked. We mentioned this
- 9:46briefly, but the golden rule here is keeping
- 9:49roles completely separate. You have to avoid
- 9:51functional overlap at all costs. The market researcher
- 9:55structures their data strictly so the strategist
- 9:58can read it. They don't try to draft the social
- 10:00posts. It's like stacking Lego blocks of data.
- 10:03Each agent perfectly builds on the previous one
- 10:06without stepping on any toes. That's exactly
- 10:08how the system avoids chaos. You use a built
- 10:11-in command called slash agents to generate these
- 10:13roles. Claude actively asks what the agent should
- 10:16do, which tools it can use, and which skills
- 10:18it has access to. And it saves a permanent agent
- 10:21file inside your project. Yeah. That file defines
- 10:24the specific working style for that digital employee.
- 10:27You keep their system prompts incredibly simple.
- 10:30You outline their main job. You list their specific
- 10:32skills. Most importantly, you explicitly list
- 10:36what they must never do. How do you stop agent
- 10:39boundary bleed when marketing tasks naturally
- 10:41overlap? Well, you clearly specify in the agent
- 10:44system prompt which upstream or downstream agent
- 10:47they must hand the work off to. So clear boundaries
- 10:51prevent the digital chefs from constantly fighting.
- 10:53Good fences make for very good AI agents. We
- 10:57have the specialized team. We have the comprehensive
- 10:59playbooks. Yep. But if you drop a massive project
- 11:03on their desk, who grabs it first? We need routing
- 11:06rules. This is the actual operating system of
- 11:08the entire agency. Without routing rules, Claude
- 11:12might call the wrong agent. Or it might use a
- 11:15heavy slow agent when a simple skill would work
- 11:17perfectly. We update the overarching manager
- 11:19file with these logic rules. Right. The simplest
- 11:21heuristic to use is this. Use a skill for simple
- 11:24execution and formatting. Use an agent for tasks
- 11:27needing judgment, deep thinking, or synthesis.
- 11:30If you just need a format change from a blog
- 11:32to an email, use a skill. Exactly. If you need
- 11:35to figure out a controversial campaign angle,
- 11:37triggers the campaign strategist agent. Let's
- 11:40look at the ultimate test outlined in the source.
- 11:43The Japan cherry blossom season campaign. This
- 11:46is where the theory gets real. Yeah. We give
- 11:48Claude one complex campaign task. Create a full
- 11:52Japan cherry blossom season campaign for GoTravel.
- 11:56We're asking for market research, a strategic
- 11:58brief, social posts, ad creatives, and a fully
- 12:02coded landing page. All from one prompt. Walk
- 12:05us through how they pass the baton. Okay. First,
- 12:07the manager reads the routing rules and activates
- 12:10the market researcher. It crawls for current
- 12:13travel trends. It discovers that travelers are
- 12:16actually exhausted by crowded Kyoto spots. They
- 12:19want hidden neighborhood gems. Next, the campaign
- 12:21strategist takes that specific data block. Right.
- 12:24It creates a strategic brief called Sakura Like
- 12:27a Local. The ankle isn't just about pretty flowers
- 12:29anymore. It's about a highly local, guided, less
- 12:33touristy experience. Wow. The brief rigidly defines
- 12:36the core message and the specific travel offer.
- 12:39Then the content creator agent steps in. Yeah.
- 12:41It's instructed to draft posts based only on
- 12:44that specific brief. It can invent new offers.
- 12:48It writes engaging posts about local neighborhood
- 12:50etiquette or common mistakes eager tourists make
- 12:53during the season. After that, the creative designer
- 12:55activates the social creative design skill. It
- 12:59references the style guide to make visuals matching
- 13:01that exact local vibe. Finally, the landing page
- 13:04skill uses all those assets to build the actual
- 13:07site. It uses the exact same message, the approved
- 13:11offer, and the foundational audience insight.
- 13:13Ooh. Imagine scaling this to a billion queries.
- 13:18Yeah. You could run hundreds of highly specialized
- 13:21localized campaigns simultaneously while you
- 13:24sleep. It's entirely possible today because every
- 13:27final output is logically connected. It's not
- 13:29a random quirky caption or a generic cherry blossom
- 13:32image. Everything flows downstream from one centralized
- 13:36strategic brief. How do you break down these
- 13:38massive complex campaigns so that the system
- 13:40doesn't crash halfway through? Well. You mandate
- 13:43in the routing rules that the AI must output
- 13:46a detailed step -by -step checklist before executing
- 13:48any actual work. So forced planning prevents
- 13:51the system from hallucinating wild ideas. It
- 13:55anchors the AI to a logical sequence before it
- 13:57writes a single word. This automated system works
- 14:00beautifully in a vacuum. Sure. But how does a
- 14:03human team actually interacts with it during
- 14:06a chaotic Tuesday afternoon. You have to connect
- 14:09it to the real world, use a Kanban board. The
- 14:11source suggests using a simple Notion task board,
- 14:14a standard layout. Yeah. To do. In progress.
- 14:17Complete. Real marketing work doesn't happen
- 14:19inside one clean, magical prompt. Tasks come
- 14:22from client emails, Slack messages, everywhere.
- 14:25Yeah. You need a centralized place where human
- 14:27managers can drop tickets for the AI to pick
- 14:30up. But you can't just write create social posts
- 14:32on an ocean ticket and expect magic. Oh, that's
- 14:34a recipe for disaster. The AI will just guess
- 14:37the context. You need highly specific bounded
- 14:39briefs. A good ticket looks like. Create five
- 14:42carousel slides for first -time Japan travelers
- 14:45based on the Sakura brief. The clearer the parameters
- 14:47of the task, the better the AI executes the workflow.
- 14:51And the source highlights a fascinating feature
- 14:53here called slash remote control. This feature
- 14:55is wild. It generates a secure temporary link
- 14:59for your active workspace. Okay. You can open
- 15:02it on your phone's browser. It lets you trigger
- 15:04your local computer's quad session from anywhere
- 15:07in the world. You could literally be buying coffee,
- 15:10check your phone, and text your AI. Yeah. Grab
- 15:14the highest priority Notion ticket and execute
- 15:16it. The AI securely reads the Notion board API.
- 15:19A digital bridge letting two different software
- 15:22programs talk. Right. It picks up the top task.
- 15:25It reads the writing rules to decide which agent
- 15:27to use. It does the heavy lifting. It saves the
- 15:30deliverables in the correct working folder. And
- 15:33finally, it marks the Notion ticket as complete.
- 15:36So what does this all mean? It means we've genuinely
- 15:39moved from querying an AI chatbot. to managing
- 15:43a digital co -worker pulling tickets from a queue.
- 15:46It's a profound shift in daily productivity.
- 15:49You aren't manually assigning tiny copy -paste
- 15:51steps anymore. You're managing an intelligence
- 15:54system. What happens if a human writes a notion
- 15:56ticket that is just too vague for the system
- 15:59to process? The AI will confidently generate
- 16:01a highly generic, useless output, proving that
- 16:05garbage in always equals garbage out. So vague
- 16:09inputs still guarantee completely useless AI
- 16:11output. Always. You absolutely cannot automate
- 16:14bad management. Let's recap the big idea here.
- 16:17Beep. The real shift isn't about writing better
- 16:21single prompts. Yeah. Prompts engineering is
- 16:24becoming obsolete. it's about building a structured
- 16:26operating system by combining permanent brand
- 16:29context rigid skills specialized agents and strict
- 16:32routing rules you create something entirely new
- 16:35you create a reliable system that can logically
- 16:38plan meticulously create and organize aligned
- 16:41marketing campaigns autonomously. It's a full
- 16:44agency living inside a code editor. It completely
- 16:46stops being a novelty toy. It becomes a functional
- 16:48business engine. Which brings me to a final,
- 16:50slightly philosophical thought. Two sec silence.
- 16:52If a solo user can build a five -person agency
- 16:55inside a code editor today. What happens in a
- 16:58few years when your autonomous AI agency starts
- 17:01independently communicating and negotiating with
- 17:04other companies' autonomous AI agencies? That's
- 17:06a wild thought. Full machine -to -machine commerce.
- 17:09Your AI negotiating a sponsorship deal with my
- 17:11AI. We'll leave you with that to ponder. For
- 17:14now, grab a notebook today. Write down just three
- 17:16repeatable weekly workflows that you can transform
- 17:19into your very first AI playbooks. Start small.
- 17:23Pick the repetitive tasks you absolutely hate
- 17:25doing. Exactly. Thank you for joining us on this
- 17:28deep dive. Stay curious. OTPO Music.