This AI Builds Your Content Team

Beyond Chatbots: The Rise of AI 'Agents'

Most people still treat AI like a smarter search box: you ask a question, it spits out an answer, and the relationship ends there. That mindset breaks down once models can not only respond, but also decide what to do next, pull in outside tools, and run for hours without supervision.

That’s where agentic AI comes in. Instead of a single prompt-response loop, an agent plans a goal, decomposes it into subtasks, calls specialized tools or APIs, and stitches the results into a finished product. Think less “chatbot” and more junior producer who can research, draft, revise, and publish while you watch the metrics.

Ethan Nelson’s Claude Code setup shows what this looks like in the wild. Built on Claude Code with multiple Model Context Protocols (MCPs), his system acts as a full-stack content team: researcher, strategist, copywriter, and project manager. He reports selling over $200,000 worth of AI systems in the last year, and this is the infrastructure behind that output.

Instead of manually prompting a model for every step, Nelson manages a network of agents. One layer hunts for trends across YouTube, Instagram, and TikTok; another analyzes outlier videos that outperform channel averages; a third reverse-engineers hooks, titles, and story structures from transcripts. The agents then feed those patterns into copywriting frameworks like AIDA and problem–agitate–solution.

On top of that, the system taps into his personal knowledge stack. MCPs connect to Notion databases holding a content intelligence vault of high-performing hooks, a Readwise-powered highlights archive from books like “Superintelligence” and “Your Next Five Moves,” and a running log of voice-note ideas. Agents pull from these sources to keep scripts grounded in his own voice rather than generic AI sludge.

Day to day, Nelson no longer “prompts” an AI; he orchestrates one. He reviews plan-mode breakdowns, tweaks workflows, and stays “human in the loop” while agents auto-generate Kanban cards, update script status, and repurpose long-form pieces into shorts. The job shifts from typing clever prompts to managing a semi-autonomous content organization that happens to live inside a code editor.

The Content Factory Blueprint

Content starts as a vague hunch: “I should talk about AI agents,” or a half-baked hook from a shower thought. Claude Code turns that into a factory line, with Claude acting as the central orchestrator that routes each step to the right specialist agent, then pulls everything back into a coherent script, newsletter, and short-form clips.

At the top sits a coordination layer. Claude asks the creator what idea they want to tackle, then fans that request out to a swarm of agents: trend research, YouTube analytics, web research, script structure, hook generation, and a strategy agent that knows the creator’s life, goals, and audience.

End-to-end, the workflow looks less like chatting with an LLM and more like running a production pipeline. A single idea can move from a one-line prompt to a long-form video script, a newsletter draft, and multiple short-form variants for Instagram, TikTok, and YouTube without the human touching a blank page.

Human control stays central by design. Nelson positions himself as the strategist and editor, not the typist: he approves the initial angle, checks the research, punches up hooks, and gives final sign-off on scripts before they move into “ready to film” status inside Notion and ClickUp.

Under that orchestrator sit three primary layers. The Research layer uses MCP-connected agents to pull trending videos, titles, and hooks from YouTube, Instagram, and TikTok, then flags outliers that overperform their channel averages.

Those research agents extract transcripts, analyze hook duration, emotional triggers, and formats (question, bold statement, contrarian take), and log everything into a “content intelligence vault.” Each entry can carry performance metrics like views and comments and a generated breakdown: “here’s why it works” and “here’s the formula.”

The Content Creation layer turns those patterns into new material. Script-structure and copywriting agents assemble outlines using frameworks like PAS (problem–agitate–solution) and AIDA (attention–interest–desire–action), while a “vibe marketing” agent enforces tone, pacing, and brand voice across long-form and short-form outputs.

A Personal Context Layer prevents the system from sounding generic. Through Notion and Readwise, Claude Code can reach into a database of highlights from “Superintelligence,” “Your Next Five Moves,” personal notes, voice-memoed ideas, and past scripts, then weave those references and opinions into each new piece so it still sounds like Nelson, not a template.

Your Automated Intelligence Swarm

Forget endless tab-hopping between YouTube Studio, TikTok search, and Notion. Ethan Nelson’s system spins up an automated intelligence swarm that does the grunt research before he writes a single line, turning hours of manual digging into a background process that quietly runs while he stays “human in the loop.”

At the front of that swarm sits a Trends Research Agent wired into Instagram, TikTok, and YouTube. It queries platforms for what’s spiking right now, pulls down trending videos, and rips out titles, hooks, formats, and durations. Instead of “what’s working this week?” being a guessing game, it becomes a live feed of pattern-rich examples.

Alongside it, a YouTube Analytics Agent hunts for statistical outliers. It scans channels, flags videos that perform significantly above a creator’s baseline, and then drills into: - View counts and growth velocity - Relative performance vs channel average - Hook length and structure - Topic, angle, and thumbnail/title pairing

Those agents don’t just surface links; they reverse engineer. Each high-performing clip gets broken into structured data: hook transcript, emotional trigger, hook type (question, bold statement, contrarian take, story), platform, and quality ranking. When comment and view data is accessible, the system scores “hook quality” based on engagement and performance metrics.

All of that flows into a Notion database Nelson calls the Content Intelligence Vault. Every entry stores why a hook works, a distilled formula, and the original performance context. Over time, this becomes a living library of proven hooks, titles, and formats sourced from across the internet, not just one creator’s intuition.

Claude Code taps that vault in real time. When a new idea comes in, agents can pull similar high-performing hooks, adapt structures, and even generate fresh variants that match a target platform and audience. Creativity starts from a stack of validated patterns instead of a blank page.

For anyone building similar swarms, Anthropic’s own guide, Claude Code: Best practices for agentic coding - Anthropic, reads almost like the spec sheet behind systems like Nelson’s. The result is a content workflow where “research” becomes a data pipeline and “gut feel” finally has numbers behind it.

Injecting Your Soul into the Machine

Soul is the bottleneck in most AI content systems. Ethan Nelson’s solution is a Personal Context Layer that hardwires his own books, notes, and half-baked ideas directly into Claude Code, so the agents never default to generic, SEO-flavored sludge.

Instead of a single “style guide” prompt, Nelson routes Claude through a mesh of Notion databases synced with Readwise. Every Kindle highlight, every offhand voice note, every framework he lives by ends up in a structured knowledge graph the agents can query in real time while drafting scripts, newsletters, and short-form video hooks.

The integration starts with Readwise’s automatic Kindle sync. Nelson pays a few dollars a month so that highlights from “China Root,” “Think on These Things,” “Do the Work,” “Superintelligence,” and “Your Next Five Moves” land in Notion with fields for book title, last highlighted date, and number of highlights.

Claude Code, via MCP connectors, can then search those thousands of quotes on demand. When a content agent needs a metaphor about discipline or uncertainty, it doesn’t hallucinate; it pulls a verbatim passage from Krishnamurti or a tactical line from a Ryan Holiday-style playbook and cites it as raw material.

Parallel to that, Nelson maintains a notes database populated by voice memos. Any time he has a shower-thought about AI agents, audience psychology, or solopreneur burnout, it becomes a discrete record. Agents mine this for recurring themes, so the system orients around what he actually obsesses over rather than what YouTube thinks should trend.

The magic shows up when the Personal Context Layer fuses sources no human researcher would casually combine. In the video, Nelson demonstrates a prompt that asks Claude to merge ideas from “The Spirit of Hope” and “Think on These Things” into a single thesis about AI and human agency.

Instead of surface-level summaries, the system cross-references specific highlights: passages about existential uncertainty from one book with meditations on attention and thought from the other. The output becomes a non-obvious claim about AI as a tool for cultivating, not numbing, awareness—something you won’t find in a stock “10 AI hacks” script.

That synthesis anchors every hook, outline, and CTA in his own intellectual history. The result: AI-generated content that sounds like Ethan Nelson because, under the hood, it literally runs on his reading habits, mental models, and late-night notes.

The Content Creation Assembly Line

Content doesn’t jump from idea to finished script in Nelson’s world. It moves through a content creation assembly line: a chain of agents, each responsible for a specific creative decision, all orchestrated by Claude Code in parallel instead of one long prompt-and-pray pass.

First in line, the Hook Generation Agent raids the Content Intelligence Vault. That Notion database tracks hooks by platform, emotional trigger, hook type (question, bold statement, contrarian take), quality score, and even duration, then pairs each record with performance metrics like views and comments when available.

When Nelson feeds it a topic, this agent doesn’t just remix a single example. It cross-references: - Top-performing YouTube long-form intros - Short-form patterns from Instagram and TikTok - Previously ranked hooks in the vault

From there, it spits out candidate hooks that already map to proven formulas: curiosity gaps, “you vs them” framing, or high-stakes promises. Each suggestion comes with a short rationale and a pattern label, so Nelson can see which psychological lever—status, fear of missing out, identity—is doing the heavy lifting.

Next up, the Script Structure Agent turns a hook and idea into a scaffold. It builds outlines with clear sections: cold open, context, story beats, value delivery, objection handling, and call to action. For YouTube, it optimizes for retention arcs; for newsletters, it biases toward narrative flow and scannable subheads.

This structure agent doesn’t hallucinate in a vacuum. It pulls from the Content Intelligence Vault’s analyzed transcripts, copying successful pacing tricks like “promise → proof → payoff” or early teaser callbacks that keep people from bouncing in the first 30 seconds. The result is a skeleton that already matches platform norms and audience expectations.

Once the outline locks, the Copywriting Agent starts filling in the meat. Nelson wired it with frameworks from Dan Co’s “two-hour writer” and classic formulas like AIDA (Attention, Interest, Desire, Action) and PAS (Problem, Agitate, Solution), plus patterns from David Ogilvy-era copywriting giants.

This agent turns bullet points into punchy paragraphs, sharp transitions, and concrete examples. It also knows Nelson’s recurring themes—AI systems, human-in-the-loop workflows, creator leverage—by pulling from his Readwise highlights, notes, and existing content database, so the script leans on his actual ideas, not generic AI filler.

Last in line, the Vibe Marketing Agent acts as the brand bouncer. It checks the finished script against Nelson’s personal context layer: preferred sentence length, level of technical depth, tolerance for jargon, and how often he references his own business numbers or client results.

If the copy drifts into robotic, over-formal, or hypey territory, this agent rewrites for tone while preserving structure and key claims. It can also re-style the same script for different surfaces—Twitter thread, newsletter, YouTube video—without losing the throughline of Nelson’s voice.

Inside the Orchestrator's Mind

Inside Ethan Nelson’s setup, Claude doesn’t just answer prompts; it runs the entire show as a workflow orchestrator. Inside Claude Code, the model reads a high-level schema of the content pipeline, then spins up specialized agents—trend research, YouTube analytics, script structure, hook generation—like a director assigning scenes to different crews.

Plan Mode is where that direction becomes visible. Before anything runs, Claude lays out a step-by-step execution map: which MCPs it will call, what data it will pull from Notion or Readwise, how it will transform transcripts into hooks, and where human review fits. Nelson can approve, edit, or kill steps, keeping a human hand on the steering wheel without doing the driving.

Those steps rely on Model Context Protocols (MCPs), which act as Claude’s I/O layer. MCPs connect Claude Code to: - YouTube analytics - Web research and live internet search - Instagram, TikTok, and YouTube trend feeds - Notion databases like the Content Intelligence Vault and Readwise highlights

Instead of hardcoding APIs into a bespoke backend, MCPs give Claude structured, permissioned access to tools and data, all from inside the coding environment.

Parallelism is where this architecture stops being a cute demo and starts feeling like a content factory. While one sub-agent dissects outlier videos on YouTube, another parses transcripts for hook patterns, and a third pulls supporting quotes from Nelson’s Readwise-powered highlights. Claude, as orchestrator, merges those outputs into a single script draft that already knows the trend, the angle, and the receipts.

Plan Mode also makes this parallelism auditable. You see which sub-agent grabbed TikTok trend data, which one ranked hooks by click-through potential, and which one injected personal anecdotes from Nelson’s notes database. If a step misfires—pulling the wrong channel, misreading a trend—you can surgically adjust that node in the plan without tearing down the whole system.

Anyone can experiment with similar orchestrated workflows at Claude.ai - Create with Claude. Nelson’s build shows what happens when agentic AI stops acting like a chatbot and starts behaving like a production team that can explain its every move before it hits publish.

The Automated Repurposing Engine

Creators don’t burn out writing; they burn out repeating themselves. A 20-minute YouTube script or 1,500-word newsletter becomes a single post on one platform, while TikTok, Shorts, Reels, X, and LinkedIn sit empty. Ethan Nelson’s Claude-powered system turns that bottleneck into a repurposing engine that treats every long-form piece as raw material for a week’s worth of distribution.

Once a flagship script lands in the content database inside Notion, Claude Code doesn’t just summarize it. The orchestrator breaks it into segments: cold open, core argument, proof, story beats, and call to action, all tagged with timestamps, emotional tone, and target audience. That structure gives the agents something to slice, not just something to paraphrase.

Dedicated repurposing agents then mine the script for: - High-signal key takeaways that stand alone in under 30 seconds - Punchy quotes and contrarian lines that read like native social posts - Narrative hooks that work as cold opens for TikTok, Reels, and Shorts

Because Claude already tracks hook patterns in the Content Intelligence Vault, it can match each segment to proven formats: question, bold statement, story, or contrarian take. A single insight about “building AI systems instead of selling prompts” might spawn five different hooks tuned to different platforms and audiences.

From there, the system generates full short-form scripts, not generic captions. Each micro-script includes a 1–3 second scroll-stopper, a tight 2–4 sentence narrative, and a platform-specific CTA, all aligned with Ethan’s Personal Context Layer so the voice still sounds like him, not a template. The same core idea can become a 45-second TikTok, a 30-second Reel, a YouTube Short, and a LinkedIn post without manual rewriting.

One long-form video that used to equal one upload now reliably explodes into 10–20 assets. Instead of asking “What do I post today?”, the only question left is “Which version of this idea goes live next?”

Blueprint for Your First AI Agent

Blueprinting your first agent starts with resisting the urge to recreate Ethan Nelson’s entire swarm. You want one narrow, end-to-end workflow that proves the concept: take a raw idea and turn it into a ready-to-publish script or post, with Claude doing most of the legwork and you staying in the loop.

Your starter toolkit looks refreshingly small. You need Claude (via API or a terminal like Warp), Notion as your structured brain, and Readwise to drip-feed highlights from Kindle, articles, and tweets into that brain. Nelson built a six-figure system on those three pillars plus a few MCPs; you can get 80% of the value with just the basics.

Notion does the heavy lifting because agents can’t reason well over chaos. Start by creating 2–3 clean databases instead of one giant doc:

• A Hook Vault: fields for Title, Hook, Platform, Emotional Trigger, Hook Type (question, bold statement, contrarian), Performance Notes, Quality Score.
• A Notes DB: Idea, Source (voice note, conversation, client call), Date, Status, Relevance Tag.
• Optional: a Readwise Highlights DB synced via Readwise’s Notion integration.

Structure matters more than volume. Nelson’s Content Intelligence Vault works because every hook includes platform, duration, emotional trigger, and a short “why it works” analysis. That schema lets Claude query “top 10 contrarian YouTube hooks with high engagement” instead of wading through unlabelled text.

Once the data layer exists, your first real “agent” is a command, not a model. Nelson calls his master prompt ID8: a single instruction that fans out into trend research, hook generation, script drafting, and repurposing. You want a similar trigger that encodes your workflow step-by-step.

A solid starter command might say: “When I give you an idea, pull 5–10 relevant hooks from the Hook Vault, combine them with 3–5 of my Readwise highlights, propose 3 angles, then draft a 2-minute script with an AIDA hook and 3 TikTok cut-downs.” Claude then uses tools or API calls to hit Notion, filter rows, and assemble the pieces.

Treat that command as code. Version it, tighten it, and bolt on new steps as you add agents, instead of rewriting your entire system every time you get a new idea.

The Human-in-the-Loop Advantage

Human creators do not disappear in Ethan Nelson’s Claude Code factory; they move up the org chart. Instead of cranking out hooks and thumbnails, Nelson acts as editor-in-chief, deciding which ideas matter, which trends to chase, and which scripts actually ship.

He stays “human in the loop” by design. Claude proposes topics pulled from YouTube, Instagram, and TikTok outliers; Nelson chooses which ones match his thesis on AI, business, and systems, then vetoes anything that feels off-brand or opportunistic.

Think of the system as a content studio where the AI swarm handles production, but the human runs strategy. Nelson sets constraints like “no scammy income claims,” “anchor in books I’ve actually read,” and “speak to builders, not spectators,” then Claude’s agents execute against those rules across scripts, newsletters, and short-form clips.

That stands in sharp contrast to fire-and-forget tools that spit out a 1,000-word blog post from a single prompt. Those systems rarely touch a creator’s Readwise highlights, Notion strategy docs, or performance data, so the output feels like reheated LinkedIn advice, not a specific person thinking in public.

Nelson’s workflow shows why full automation actually weakens content. When agents mine his Readwise database, content intelligence vault, and personal notes, they surface raw material; he still decides which quote from “Superintelligence” supports a point and which trend hook crosses into clickbait.

True leverage comes from augmentation, not abdication. Claude Code handles: - Trend scouting and transcript scraping - Hook pattern analysis and ranking - First-draft scripting and repurposing

Nelson handles: - Narrative angle and positioning - Ethical and factual guardrails - Final line edits and on-camera delivery

Developers can wire up similar human-centered pipelines using tools like the Anthropic Python SDK - GitHub, then plug into Notion, Readwise, or custom analytics. The result is not an AI that replaces the creator, but a specialized team that makes their taste, judgment, and voice scale across every channel they care about.

The Future is a Co-op: Working With Your AI

Future creative work looks less like typing into a single chat box and more like directing a small studio of AI agents. Managing half a dozen specialized workflows—research, scripting, repurposing, analytics—will sit next to filming and editing as a baseline skill, the way knowing your way around a CMS or Premiere does today.

Creators already running systems like Ethan Nelson’s Claude Code setup are effectively early adopters of this role. They’re not “using AI” so much as assigning work, reviewing outputs, and tuning agents the way a showrunner steers a writers’ room.

Next-generation agents won’t stop at drafting scripts and carousels. They will: - Autonomously schedule posts across YouTube, TikTok, Instagram, X, and newsletters - Analyze comment sentiment and watch-time curves in real time - Propose new content angles based on what converts followers to buyers, not just what gets views

Tie that into ad platforms and CRM data, and your system starts testing thumbnails, hooks, and offers automatically, then reallocating effort toward what moves revenue. Think of an agent that notices your “AI workflows” shorts have a 40% higher save rate and quietly spins up a weekly series, complete with outlines and filming checklists.

The job title quietly morphs from “prompt engineer” into AI systems manager. You won’t obsess over crafting a perfect single prompt; you’ll design policies, guardrails, and handoffs between 10–20 agents, then audit their logs the way an SRE watches dashboards.

You’ll decide which data each agent can touch—analytics, personal notes, client docs—and how they escalate edge cases back to you. Your value shifts toward judgment: what to automate, what to kill, and where your taste and experience change the output.

Start by treating your workflow as a system, not a pile of tasks. Map how ideas move from spark to script to multi-platform distribution, then highlight every repetitive step that doesn’t require your taste, face, or voice.

Those boxes are your first agents. Build one, wire it into your tools, and keep going until your default question changes from “Can AI do this?” to “Which part of my system should own this next?”

Frequently Asked Questions

What are Claude Agents in this context?

They are specialized AI processes orchestrated by a central Claude instance to perform discrete tasks like research, writing, and data analysis as part of a larger workflow.

How does this system maintain an authentic voice?

By integrating with the creator's personal knowledge bases in Notion, including book highlights from Readwise, personal notes, and unique ideas, the AI learns to write from their specific perspective.

Is this a specific Claude feature or a custom build?

This is a custom-built workflow that leverages Claude's core capabilities within a terminal or API environment. It connects to external tools and databases to create a unique, automated system.

What is the 'Content Intelligence Vault'?

It's a custom database in Notion that stores and analyzes high-performing content hooks, titles, and formulas that the AI finds online, creating a repository of proven ideas for new content.