In This Conversation

• The six lanes, and which four are actually covered. Publishing, notes, inbox replies, analytics, restacks and reshares, DMs and notifications. Four are automatable today. Two are not, and I say which.

• Substack’s own bestseller MCP is read-only with a handful of data points, which is why this exists at all. About 20 of the 44 tools are free, and the reverse-engineering write-up is public so you can build your own instead.

• Retroactive tables of contents across an entire archive, with Kim pointing out the SEO upside and me arguing against doing it all at once.

• MCP against CLI, and why I think the trend is moving toward CLIs even though this whole demo is an MCP.

Watch the replay above. hosted and kept interrupting with the questions I would have skipped, which is what makes this one worth the hour. Every demo ran against my live Substack account.

Key Moments

• 00:26: Kim opens and sets the frame: using MCPs with Claude and ChatGPT to run Substack

• 01:14: The problem. The more I work in Claude, the more control I have, and Substack is the one place that connection was missing

• 02:00: Substack's own bestseller MCP is read-only with only a few data points

• 02:30: The six lanes laid out. Newsletter, notes, engagement, analytics, restacks and reshares, inbox

• 06:22: Kim asks the room whether growth feels slow right now, and nobody disagrees

• 08:00: Why a scheduler alone stopped being enough. Even with one, it is still click-ops

• 10:00: Why Cowork rather than chat. You can build an agent, and an agent can run on a schedule

• 11:00: The social-proof workflow: drop a screenshot in a folder, one note goes out at 9:30 the next morning

• 13:33: Kim asks why Sonnet. Because the workflow is simple, and the complexity was already solved upstream

• 14:30: Connecting live. Add custom connector, paste the URL, then either a magic link or a pairing code

• 18:00: Asking it for calendar capacity for the next seven days

• 19:30: Five notes from the latest post, scheduled. Then the version that walks the last 30 posts

• 22:04: Kim's question: can you force a voice skill or a copy rubric onto it? Yes, and two ways

• 24:00: Retro-checking already-scheduled notes against your brand voice and rescheduling the ones that miss

• 27:00: Viral note formats, drawn from the mined playbook, all of it in the free pack

• 29:30: High-CTR title options for posts that underperformed, including ones from six months ago

• 32:00: Tables of contents. Fetch the latest post, check if one exists, build it, publish

• 34:00: Kim: doing it retroactively may earn a little juice because Substack sees the update

• 34:30: My counter: do not mass update. Schedule an agent to do five or seven a day. Mass updates are a signal

• 36:30: The honest answer. Did this grow the publication? I do not know. Month over month is declining

• 39:00: Kim asks about scanning old posts for missing internal links, and why that one is genuinely harder

• 42:00: Analytics: comparing all published posts and asking what the best openers have in common

• 44:00: The engagement inbox, drafting warm replies, and where I refuse to automate a real relationship

• 46:28: Back to the six lanes. Four covered, restacks and reshares plus DMs still open

• 50:07: Kim publishes to her own website first and does not trust platforms, so can an MCP push from there? Yes

• 51:19: Kim's thesis: MCPs are becoming a new kind of digital product

• 52:04: Pulling other people's publications into a feed, and why that sits in a grey area of the terms

• 55:12: MCP against CLI. Tavily, Notion and Gumroad have all shipped CLIs, and I think that is the direction

• 57:34: The plain-English difference: a CLI is a binary on your machine, an MCP is a network call out

• 1:01:31: Why CLIs use fewer tokens. An MCP carries contracts, tool definitions, schemas and descriptions

Top Takeaways

1. Four of six lanes, and be honest about the other two. Publishing, notes, analytics and inbox replies are covered. Restacks, reshares, DMs and notifications are not. A tool that claims all six is overclaiming.

2. Never mass-update your archive. Kim is right that Substack notices an update, and that is exactly why a sweep across 90 posts in one afternoon reads as a mass edit. Five to seven a day on a schedule.

3. The brand voice can run as a gate, not just a generator. Ask it to fetch the next seven scheduled notes, check them against your voice skill, and reschedule the ones that miss. That is the step most people skip.

4. I cannot prove this grew anything. Month over month is down, and it may be summer, or too many emails, or too many Lives. What I can say is I stopped spending the hours, and those hours went into the cohort and YouTube instead.

5. The trend is running toward CLIs. An MCP carries contracts, tool definitions, schemas and descriptions on every call. A CLI is a binary on your machine. Tavily, Notion and Gumroad have all shipped one.

6. Do not scrape other people's publications. Pulling your own activity feed is fine. Building a public tool that reads other writers' content sits in a grey area of the terms, and I will not ship that.

Learn with me: Claude Code Builder cohort

I run the Claude Code Builder cohort on Maven: four weeks, six live sessions, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Build it with me, live. See the syllabus →

Resources Mentioned

• Claude Cowork: where the agent lives, because a chat cannot run on a schedule.

• Finn Tropy: pointed out in the chat that the inbox has its own APIs and that pulling your own activity feed should be fine.

• Kim Doyal: host of the Live, publishes to her own site first on principle.

• PluggedIn MCP: the announcement and the endpoint at substackmcp.genaiunplugged.com, roughly 20 of 44 tools free to subscribers.

• Reverse-engineering Substack's API: the write-up for building your own server rather than using mine.

• Scheduling notes with Claude Cowork: the free walkthrough Kim asked me to name on air.

• SubflowAI: the Chrome extension Kim already uses, and the data layer behind the analytics answers.

Where This Goes Next

Connecting the tool is the easy half. The part that takes real time is the layer underneath: the agents that decide what to write, the memory they read, and the loop that stops their voice drifting back to generic.

PluggedIn subscribers get the working versions of what I demo on these Lives: the agent files and prompts you can drop into a folder tonight, the workflow bundles, and the pipeline pieces I run on this publication every week.

Connect

• Kim Doyal: on Substack

• Dheeraj: GenAI Unplugged | YouTube

Which of the six lanes would you hand over first, and which one would you never hand over? Tell me in the comments.

I sat down with , a fellow solo creator who runs his own newsletter for other solopreneurs, to solve something he'd been asking me about for a month.

He wanted his Substack connected directly to Claude, not through another dashboard, but as an actual agent that could read his posts, write his notes, and touch his publication the way he would, just faster.

By the end of our session, we had it working live. Claude read Anfernee's calendar, rewrote an old post into five scheduled Notes, and inserted a table of contents into an already-published article, all without re-emailing a single subscriber.

Solopreneur Code

Weekly ideas, systems, and AI workflows for solopreneurs who want to build a profitable one person business without the hustle. Subscribe and get the free Solopreneur Success Hub.

By Anfernee

This article is the walkthrough for anyone who wants the same setup. I'll show you what the four-job grind looks like, why a nice dashboard doesn't fix it, and the exact steps to wire an MCP connector into Claude so it can run your Substack operations from a single sentence.

MCP stands for Model Context Protocol, the standard that lets an AI assistant call real tools on a real platform.

"It usually takes me about half an hour to just do one." - Anfernee, describing his old process for building a table of contents by hand

Why Does Running a Substack Require So Much Manual Work?

Substack isn't just a newsletter tool. It behaves like a social network layered on a publishing platform, so you're maintaining four manual tracks at once: writing the newsletter, posting Notes daily, engaging with other posts, and restacking content to stay visible in the feed.

Skip any one of those tracks for a week and your reach quietly drops. I've watched it happen on my own account.

The guidance I keep coming back to is posting a Note roughly every day, on top of everything else. That's exactly the kind of task that never gets easier just because you've done it a hundred times.

I built my first fix for this problem, Subflow AI, because I got tired of clicking the same sequence of buttons every morning.

It started as a Chrome extension scheduler, a personal tool, because Substack had no native scheduling and the other options on the market felt expensive for what they did.

I was already running an n8n workflow to patch the gap. It worked, but it cost money and still needed babysitting.

Anfernee had lived the earlier, worse version of this problem directly.

"I recall like one month ago, I just WhatsApp Dheeraj and said, hey, do you have an MCP connector?" - Anfernee

Before Substack added native scheduling, he physically couldn't post during the hours he was asleep. That's not a metaphor. If he wanted a Note live at 7am, he had to be awake to click publish at 7am.

Once scheduling arrived, and later once Subflow AI gave him a calendar view of what was queued, that constraint disappeared. But even with a good dashboard, he was still the one doing every click.

Why Doesn't a Good Dashboard Solve the Problem?

A dashboard, even a well-built one, only reduces friction. It doesn't remove the requirement that a human sits down and operates it, screen by screen, task by task, every single day, no matter how clean the interface looks or how many shortcuts it adds.

I learned this the hard way after shipping Subflow AI. Building the tool was maybe 20% of the work. Everything that came after, support requests, feature asks, watching how real people used it, turned out to be the harder 80%.

"That's why I always say building is just 20%. What lies after shipping is the hardest 80% of the job." - me, on the real cost of maintaining a product

Anfernee's request for an MCP connector was a direct result of that 80%, a real user hitting a real ceiling on what a dashboard, no matter how clean, could do for him.

Read more on how I built GenAI Unplugged's own MCP server if you want the deeper model behind this pattern.

The shift that fixed it wasn't a better UI. It was connecting Substack to Claude as an agent, so instead of clicking through the notes scheduler, the content repurposer, and the analytics tab separately, I could describe the whole workflow in one sentence and let Claude execute it.

Think of the difference this way: a dashboard is a car with better pedals. An MCP connection is cruise control. You're still driving somewhere, but you stop having to manage every single input along the way.

How Do You Connect a Substack MCP Server to Claude?

You connect a Substack MCP server to Claude by adding it as a custom connector using its URL, then completing pairing with a one-time code generated inside Subflow AI. Once paired, you set permission levels per tool so read-only actions run automatically while anything destructive waits for your approval.

Here's the full sequence I walked Anfernee through live, in order:

1. Add the custom connector. In Claude, open connector settings and add a custom connector using the Substack MCP's URL rather than searching a pre-built directory.

2. Generate a pairing code. Inside Subflow AI (or as a free Substack subscriber using the free tier), generate a one-time code.

3. Complete pairing. Paste the code into Claude to finish linking your Substack account to the connector.

4. Set tool permissions. Go through each individual tool the connector exposes and choose always-allow, always-ask, or deny.

5. Test with a read-only prompt. Start with something low-risk, like asking Claude to list your last five posts, to confirm the connection is live before trusting it with writes.

The permission step matters more than it sounds like it should. MCP lets you scope trust at the individual tool level, not just at the connection level. You can let Claude freely read your analytics while still requiring your explicit approval before it deletes a note or publishes a post.

Free tier access

• Best for: Free Substack subscribers wanting to test read/write tools

• Setup: Pairing code, no payment required

• Limitation: Core notes and post tools only, no agent-mode scheduling

Subflow AI paid tier

• Best for: Creators who want the full repurposing and scheduling engine

• Setup: Same pairing flow, active automatically on a paid account

• Limitation: None on the connector itself, tied to Subflow AI's existing pricing

Both free and paid users can connect the same way. The difference is what happens after the connector is live, not how you set it up.

Troubleshooting the Connection

• Pairing code expired. One-time codes are short-lived. If Claude rejects it, generate a fresh one from Subflow AI (or the free tier) and paste it in immediately.

• A tool keeps asking for approval you already granted. Check that you set the permission on the specific tool, not just the connector as a whole. Permissions are scoped per tool, not global.

• The connector doesn't show up in Claude. Confirm you used the "add custom connector" path with the direct URL instead of searching Claude's built-in directory. Custom connectors don't appear in search results.

Learn with me: Claude Code Builder cohort

I run the Claude Code Builder cohort on Maven: four weeks, six live sessions, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Build it with me, live. See the syllabus →

How Do You Automate Notes at Scale With Natural Language?

You automate Notes by asking Claude, once connected, to pull an existing post and rewrite it into multiple Notes scheduled across specific days, all inside the chat, with zero clicks in Substack or Subflow AI itself, the same way you'd type it out for a human assistant.

During our live demo, I asked Claude to take my latest published post and generate five Notes from it, spread across the coming week:

Write five Notes from my latest post and save them as drafts.
Spread them across next week, one a day, and show me the
proposed schedule before you commit to it.

It pulled the post through the connector, drafted variations against the dozen or so viral Substack note formats I've studied and fed into the system, and scheduled each one directly. What used to be a string of individual scheduling clicks became a single instruction.

See three more Claude Cowork automations if you want to take this pattern further.

That's the mental shift the whole demo turns on. Once you can name a workflow that only exists in your head, you can hand the whole thing to an agent instead of clicking through it yourself.

"The whole workflow that you have in your head, you can define it in an agent, and you can schedule that agent every week." - me, describing the mental shift this creates

You can scale this pattern further than a single post. Ask the connected agent to look at your last several published posts and generate Notes from all of them at once, with an instruction to avoid repeating the same angle twice.

That turns a week of note-writing into one prompt you run on a Sunday and forget about.

*I get asked constantly why solo creators keep rebuilding the same manual Substack checklist every week instead of just describing it once.

The honest answer is that most people don't know MCP connectors exist yet, and building one from scratch, permission scoping included, easily eats a full weekend if you're doing it without a reference.

That's the exact gap PluggedIn members skip past, since I hand them the working setup instead of the research.*

How Do You Add a Table of Contents to a Substack Post Automatically?

You add a table of contents automatically by asking the connected agent to fetch your last published post, insert the table of contents at a specified location, and republish it, all without triggering a duplicate email to your subscribers, a job that used to cost 30 minutes.

I ran this live on an already-published post of mine. Substack still has no native table of contents feature as of July 2026, so until this connector existed, building one meant manually re-reading the post, writing out anchor links by hand, and re-editing the live version.

Anfernee's number for how long that took him on his longer posts: about half an hour, every single time.

"It usually takes me about half an hour to just do one." - Anfernee

With the connector live, I asked Claude:

Can you create the Table of Contents for my last published
post and update it? Don't trigger a new email to subscribers.

It found the post through the connector, generated the structure, and pushed the update. The post updated in place, something like this dropped in above the intro:

## Table of Contents
- Why This Matters
- The 3-Step Setup
- Common Mistakes to Avoid

No re-send email hit subscribers' inboxes.

The part worth calling out here is that you can give the agent standing instructions that persist across future runs.

Tell it once where you want the table of contents placed relative to your intro, and every future post you ask it to update follows the same rule without you repeating yourself.

I use this same standing-instruction pattern for a chunk of my own PluggedIn workflows: set the rule once, let the agent apply it every time after.

What Else Can a Connected Substack Agent Do Besides Notes and Posts?

A connected Substack agent can pull real analytics, like which posts get the most opens and where subscribers are coming from, and feed that live data directly into Claude's artifacts instead of you typing numbers in by hand, turning a static Claude artifact into one that updates itself.

This was the moment Anfernee called out as the biggest personal change of the whole session.

"This is the number one thing that I couldn't do on Claude before. I had to key the numbers myself, and now I can finally build my own live artifact." - Anfernee

Before the connector, building any kind of live dashboard or chart inside Claude meant manually typing your own subscriber and open-rate numbers into the prompt every time you wanted an updated view.

Now the agent pulls current numbers straight from Substack, so the artifact reflects what's happening in your publication, not a snapshot from whenever you last copied it over. I walk through a related build in a Claude Cowork agent that plans content overnight, using the same live-data pattern.

We also tested two more prompts live:

Compare all my published posts and tell me the three best
by opens, and what they have in common.

Draft a warm reply to the first comment in my inbox, but
don't send it.

The first generates high-CTR title variants and analytics comparisons for anything you've already published. The second drafts replies to inbox comments and mentions, but only as a draft. I've kept auto-send off on purpose.

Replying is the last part of this job that still feels like it should stay human, and I'm not ready to hand that off.

Anfernee started to say it out loud mid-demo: "So if replies are also generated by AI, then, you know, it's..." I finished the thought before he could: "It's the end, right? AI talking to AI, and we're just watching."

If you want to take the automation further once you've got the connector working, my Claude Code SEO and AEO agent build is the natural next thing to wire up.

Get PluggedIn

Stop being the one person who has to remember every Substack task, when an agent can hold that checklist for you.

See what's included →

Building this connector from scratch, permission scoping included, costs a full weekend if you're doing it without a reference. Anfernee's old table of contents process alone cost him roughly 30 minutes of manual work per long post, every time.

That's the tax an unconnected Substack workflow charges you, week after week, for tasks an agent now runs in one prompt.

Get PluggedIn to go from clicking through five separate Substack screens to describing your whole weekly workflow in one sentence.

Frequently Asked Questions

What is Subflow AI?

Subflow AI is a scheduling and content-repurposing dashboard I originally built for myself as a Chrome extension, back when Substack had no native scheduling. It grew into a fuller product with a notes scheduler, content repurposer, brand-voice configuration, and analytics after other users asked me to release it.

How do you connect Substack to Claude?

Add a custom MCP connector in Claude using the Substack MCP's URL, then complete pairing with a one-time code generated inside Subflow AI or through the free tier. From there you set per-tool permissions before letting the agent run anything on your behalf.

Do you need a paid subscription to use the Substack MCP?

No. Free Substack subscribers can use the free read and write tools through the connector. Paid Subflow AI subscribers also get the deeper agent-mode scheduling and repurposing features.

Can this replace ChatGPT instead of Claude?

Yes. Any MCP-compliant client can connect to the same Substack MCP server, and Claude, ChatGPT, and Codex were all named as compatible options during our session.

How do you add a table of contents to a Substack post?

Ask the connected agent to fetch your last published post, insert a table of contents at your preferred location, and republish it. The republish does not trigger a duplicate email to subscribers.

Can you control what the AI is allowed to do automatically?

Yes. Every connected tool can be set individually to always allow, always ask for approval, or deny outright. That lets you automate low-risk reads while keeping a human check on anything destructive, like deleting a note or publishing a post.

Key Takeaways

• A dashboard reduces clicks, an agent removes the requirement to click at all. Subflow AI's UI helped, but the real shift came from connecting it to Claude as an agent.

• Permission scoping is what makes automation safe to trust daily. Set read-only tools to always-allow and destructive actions to always-ask, at the individual tool level.

• Notes at scale start with one prompt, not one click at a time. Ask the agent to repurpose your last several posts into Notes, spread across days, avoiding repeated topics.

• A table of contents that used to cost 30 minutes now costs one instruction. Give the agent a standing placement rule once and it applies it to every future post.

• Live artifacts finally reflect real numbers. Analytics pulled through the connector replace hand-typed figures in any Claude artifact you build.

• Reply automation stays draft-only by design. Technically possible to auto-send, kept as a human-reviewed step on purpose.

• The connector isn't locked to Claude. Any MCP-compliant client, including ChatGPT and Codex, can use the same setup.

Resources Mentioned

• Subflow AI: My own scheduling and content-repurposing dashboard for Substack, the product underlying this MCP integration.

• Substack MCP: The MCP server exposing Substack's notes, posts, analytics, and engagement tools to any compliant AI client.

• Claude / Claude Cowork: The AI assistant I connected the Substack MCP into for this live demo, run in Cowork mode for the scheduling and repurposing tasks.

• Claude Code: The path I'd point you to if you want to build your own version of this MCP integration from scratch instead of using Subflow AI's.

• ChatGPT: Named as another MCP-compliant client that can connect to the same Substack MCP server.

• Codex: Named alongside Claude and ChatGPT as a third compatible MCP client.

• n8n: The workflow tool I ran before building Subflow AI myself, cited as costly enough that building my own scheduler made more sense.

• One Shot Show: My Substack Live series, co-hosted with Wyndo, referenced for how advance scheduling affects attendance.

Your 15-Minute Challenge

Pick your single most repeated Substack task this week, likely turning a post into Notes or checking which post got the most opens. Set up the Substack MCP connector in Claude following the five steps above, starting with a read-only test prompt before granting any write permissions.

Once connected, ask it to do that one repeated task exactly once. If it saves you the clicking you'd normally do by hand, you've proven the setup works, and you can add the next task next week.

The problem Timo ended with: his tools don’t talk to each other. He is the one doing all the connecting. That costs 10-15 minutes of copy-paste friction per article, and produces output that varies in tone because each project carries its own context.

Write Your Way To Wealth

The exact system I use to build a $2K/mo Substack while working a 9–5. Get clear, practical strategies to grow to 1,000 subscribers, monetize your audience, and build a newsletter that compounds into real income.

By Timo Mason🤠

This article is the fix. By the end, you’ll have a clear blueprint for collapsing Timo’s 8-step workflow into one Claude Cowork agent using a 4-layer architecture, including the exact files to create and the order to build them.

Part 1 Recap: What Timo’s 8-Step Workflow Actually Looks Like

Timo covers this in full in Part 1. Here is the short version.

Eight steps, five tools: Claude Article Architect project for outlining and drafting section by section. A Substack Internal Linker skill connected to his Notion article database for internal link suggestions.

A Product CTA Linker that finds placement spots and auto-generates UTM-tracked product links. Headline Hero for title options. A separate ChatGPT GPT for Substack Notes. Another GPT for thumbnails.

Each step works well on its own. Timo was honest about the gap at 2:46 in the recording:

“I’m practical and not over-optimizing and just do the copy-pasting for two seconds.” - Timo

The copying is not what is slow. What is slow is the context reset. Every tool switch means re-explaining your article from scratch.

Every project change means losing the accumulated understanding the previous tool had built up. Multiply 10-15 minutes per article by five publications per month and you have lost over an hour. Not to writing. To tool management.

That is exactly what the agent removes.

Why Claude Cowork Is Different From Claude Projects

Claude Cowork is the agentic upgrade to Claude Projects. Projects give you a persistent workspace with a system prompt and knowledge base.

Cowork adds multi-step workflow automation, external integrations via MCP servers (like Notion and Substack), reusable skill modules shared across all projects, and scheduled task execution triggerable from mobile.

As of May 2026, Cowork is Claude’s answer to the orchestration problem that Projects leaves to you.

Here’s what that difference looks like in practice:

Claude Projects

• Best for: Single-step writing tasks, document analysis

• Setup: Create project, add files, start chatting

• Skill support: Skills exist inside the project only

• External integrations: None native

• Runs without you: No

Claude Cowork

• Best for: Multi-step workflows, newsletter production pipelines

• Setup: Select local folder, connect MCPs, write Claude.md

• Skill support: Skills are reusable modules visible across projects

• External integrations: Notion, Tavily, custom MCPs

• Runs without you: Yes, via scheduled tasks + Dispatch

Timo’s skills (Internal Linker, CTA Linker) currently live inside his Article Architect project and can’t be called from anywhere else.

In Cowork, they become reusable skill modules in Claude: any project can call them, and the unified agent orchestrates them in sequence without manual triggering.

At 49:30 in the recording, I walked through a live Cowork setup with Timo, creating a project from a local folder in a few clicks.

Learn with me: the Claude Code cohort

I run the Claude Code cohort on Maven: six live sessions across six weeks, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Use code GENAI20 for 20% off. Check the Syllabus →

CLAIM 20% OFF

Build it with me, live. See the syllabus →

How Does the 4-Layer Agent Architecture Work?

The 4-layer architecture organizes a content agent into four components: Integration (external tool connections), Memory (learned corrections over time), Context (brand identity files that inform every output), and Instructions (the Claude.md file that orchestrates each step).

Build them in this order. Each layer depends on the one below it. The full setup takes three to five hours the first time, and it compounds with every article after.

In Part 1, Timo described three of these layers: Integration, Context, and Instructions.

Memory is the fourth. It is what makes the agent improve over time without re-explanation, and most people skip it initially and add it once the other three are running. Here is what each layer contains.

Layer 1: Integration (the connectors)

This is where your external tools connect. For a Substack creator, the minimum useful set is Notion (for your article database and content calendar) and the Substack MCP (for direct publishing).

Research tools like Tavily live here if you have a research phase. If you don’t, like Timo, you skip them. The integration layer is what separates an AI assistant from an AI agent (software that acts on your behalf without step-by-step prompting). An assistant waits for you to paste things in. An agent fetches what it needs.

Layer 2: Memory and Knowledge

The memory layer is where the agent learns from your edits over time. When you consistently change the same thing, whether it’s a tonal pattern it gets wrong, a format you always adjust, or a product name it misspells, the memory layer captures that and applies it automatically to future outputs.

This layer is optional at first, but it compounds. Every correction you make becomes a preference the agent applies without re-explanation.

Layer 3: Context (the hardest part)

Skip this layer and every AI draft sounds like it could belong to any newsletter.

The context layer is a set of files that give the agent specific direction before it writes a single word.

At minimum, you need three files: an About Me file covering your background, brand, and writing voice; a content strategy file with your pillars, cadence, and quarterly goals; and a target persona file describing your reader in specific terms (not “entrepreneurs” but the actual person who subscribes and why).

At 1:04:15, Timo articulated the value better than I had:

“Not just that we save time. If we have the context files very nicely set up, that very much potentially will be a better output.” - Timo

That reframe matters. The unified agent isn’t just faster than the fragmented workflow. It’s better, because every step has access to the full context instead of just what you remembered to paste in.

Layer 4: Instructions (the Claude.md file)

The Claude.md file is the agent’s operating document. It’s a structured plain-English file that the agent reads at session start, and it tells the agent exactly what to do, in what order, calling which skills, referencing which context files. Nothing in this file is technical.

At 38:00, I told Timo: dictate your existing workflow into Whisperflow or Mac dictation, then ask Claude to structure it into a Claude.md file. The more detail you include, the better the agent performs.

A Claude.md for Timo’s workflow would look roughly like this:

# Content Agent. Operating Instructions

## Startup Protocol
On session start:
1. Read about-me.md
2. Read content-strategy.md
3. Read target-persona.md
4. Display: “Context loaded. Ready to begin. What are we working on today?”

## Article Workflow
Step 1. Outline: Ask for topic + angle. Generate outline using [outline-template.md].
Step 2. Draft: Write section by section. Match tone in [example-article.md].
Step 3. Internal links: Run Substack Internal Linker skill against Notion article database.
Step 4. CTA placement: Run Product CTA Linker skill. Auto-generate UTM links.
Step 5. Headlines: Run Headline Hero skill. Return 5 options.
Step 6. Notes: Draft 3 Substack Note variations from final article.
Step 7. Publish: Use Substack MCP to schedule Note and draft post.

## Output Format
- Article: Google Doc export format
- Headlines: Numbered list with hook score
- Notes: 3 variations (short/medium/long)

That’s it. The entire workflow is documented. The agent executes it without you managing the handoffs.

What Is the Context Layer and What Files Do You Actually Create?

The context layer is where most creators stall. Not because it’s technically hard, but because it forces you to articulate things you’ve kept implicit: who you are, who your reader is, and what you’re actually trying to build.

At 22:01, I walked through the specific files. Here’s what each one contains:

About Me file (about-me.md)

Your background, your brand name, your writing voice principles, your non-negotiables. For Timo, this would include his “good content over SEO” philosophy, his personal-experience-only content approach, and his conversational writing style. One to two pages. Not a bio, a working document.

Here’s a short example of what a completed About Me file looks like:

# About Me

Name: Timo. Publication: Write Your Way to Wealth.
Audience: Newsletter creators who want to grow and monetize.
Voice: Direct, practical, no-fluff. I write from personal experience only.
Philosophy: Good content is the best SEO. I don’t optimize. I publish.
Non-negotiables: No listicle clickbait. No vague AI hype. Real results or silence.
Products: [Paid newsletter tiers]. CTA style: honest offer, specific outcome.

If you’re unsure how to surface your own content DNA, extracting your newsletter’s DNA with Claude Cowork is a structured process that produces exactly these files in under an hour.

Content strategy file (content-strategy.md)

Your content pillars, your publication cadence, your products, and your quarterly goals. This gives the agent strategic context so it doesn’t suggest content that pulls against how you want to be known.

Target persona file (target-persona.md)

Your reader, specific. Not “newsletter creators” but the person who subscribed last week: what they’re struggling with, what language they use, what they want to achieve. One detailed profile beats five vague ones.

Article format templates (templates/)

These are structural scaffolds for your article types, not full examples. A how-to template has different section headers than a personal story or a listicle. Templates tell the agent what shape to use. A published article in your voice tells it what tone to use. Both belong in context, but they do different jobs.

One distinction worth making: a template is structure, an example is tone. Don’t confuse them.

How Does the Substack MCP Eliminate the Last Manual Step?

The Substack MCP is a custom integration server that connects Claude directly to Substack’s publishing system.

Once configured in Claude Cowork, your content agent can create, schedule, and publish Substack Notes and post drafts without leaving the Claude interface, eliminating the final copy-paste step in the workflow.

I launched this MCP the day before recording this session. At 54:03, I walked Timo through what it actually provides:

• Create, publish, edit, list, and delete Substack Notes

• Manage post drafts

• Access subscriber stats and dashboard data

For Timo’s workflow specifically: the final manual step, copying polished content from Claude into Substack, disappears.

The agent writes the Note, schedules it, and publishes it without leaving the Claude interface. His separate GPT for Substack Notes becomes unnecessary. The entire pipeline, from outline trigger to published note, runs inside one unified agent.

You can read how the Substack MCP server was built for the technical background. The MCP is available at substackmcp.genaiunplugged.com and is free for all GenAI Unplugged subscribers.

How Long Does This Take to Set Up?

The context layer takes two to four hours on the first pass. Writing your About Me file, content strategy, and target persona means articulating things you’ve kept implicit: who you write for, what your content stands for, what makes your voice specific. This is the work that determines output quality for every article after.

Once those files exist, the rest moves quickly. Connecting Notion takes ten minutes. Writing your Claude.md takes thirty minutes (especially if you dictate it first and have Claude structure it).

Registering your existing skills as Cowork modules takes another fifteen minutes. The Substack MCP setup takes under ten minutes.

Total first-time setup: three to five hours. After that, adding a new skill or adjusting your workflow is a one-line change to Claude.md.

“The context layer is the hardest part. And once that is there, everything is like more or less just kind of copy-pasting.” - Timo (at 1:05:35)

You pay the setup cost once. Every article after that runs on the foundation you’ve already built.

Common Setup Issues

Notion database not found.

Cowork connects to your Notion workspace but only sees databases you’ve explicitly shared with the integration. Open Notion, go to Settings, then Connections, find your Cowork integration, and share each database individually.

Substack MCP authentication fails.

The MCP uses your Substack session cookie for auth. If it fails, clear your Substack session, log back in, and re-run the MCP setup at substackmcp.genaiunplugged.com. Safari users: copy the cookie value from developer tools rather than using the browser extension.

Claude.md instructions are ignored.

If your agent isn’t following the startup protocol, check that the Claude.md file sits at the root level of your Cowork project folder, not inside a subfolder. The agent looks for it specifically at the top level.

Skill not recognized during workflow.

If a registered skill isn’t being called, confirm it’s listed in your Cowork skill library (Settings, then Skills) and referenced by its exact registered name in your Claude.md.

If you’re spending 10-15 minutes per article on copy-paste handoffs between separate Claude projects, that’s 60-90 minutes a month for a twice-weekly publication. Time that compounds into hours when you account for the re-explanation overhead every time you switch contexts.

Get PluggedIn

If you’re building AI content workflows on Substack, the unified agent I walked through with Timo is the starting point. The PluggedIn tier covers the full setup: the exact context files, Claude.md templates, and MCP configurations I use.

See what’s included →

Every month without the context layer costs you 60-90 minutes in manual handoffs. Time that resets from scratch every time you switch projects.

FAQ

What is the difference between Claude Cowork and Claude Projects?

Claude Projects gives you a persistent workspace with a knowledge base and a system prompt. Claude Cowork adds an agentic layer on top: you can connect MCP servers (like Notion and Substack), use skills as reusable modules visible across all projects, schedule automated tasks, and trigger execution from mobile via Dispatch.

Projects are for single-step tasks where you manage each handoff. Cowork is for multi-step workflows that run without you directing every step.

What files do I need to create for the context layer?

At minimum, three files: an About Me file (your background, brand, voice principles), a content strategy file (pillars, cadence, products, quarterly goals), and a target persona file (specific reader profile).

Optionally add article format templates for each content type you write, and one published article as a tone reference. These files become the knowledge base your agent draws from before writing anything.

Do I need coding skills to build this content agent?

No. The setup involves selecting a local folder in Claude Cowork, writing your workflow in plain English in a Claude.md file, and connecting the integrations you already use (like Notion). The recommended approach: use Whisperflow or Mac dictation to narrate your entire workflow out loud, then ask Claude to structure that transcript into a Claude.md file. The agent writes the instructions for itself.

What is the Substack MCP and how does it work with Claude Cowork?

The Substack MCP is a custom MCP server (substackmcp.genaiunplugged.com) that connects Claude directly to Substack. It provides 11 tools including creating, publishing, editing, and scheduling Substack Notes plus post management and subscriber stats. Once connected as a Cowork integration, the agent can publish directly to Substack without any copy-paste step. Available free to all GenAI Unplugged subscribers.

Can the agent automatically start writing without me triggering it?

Yes. Claude Cowork’s scheduled tasks feature lets you write an instruction telling the agent to check your Notion content calendar, identify the next article due, and start the workflow on a set schedule. Once configured, adding an article to your Notion calendar is all you need to do. The agent handles the rest.

What is a skill in Claude Cowork and how is it different from a project?

A skill is a reusable, modular prompt or procedure that any Cowork project can call as a step in a workflow. In the unified agent approach, tools like Headline Hero, Internal Linker, and CTA Linker each become skills the single agent orchestrates in sequence, instead of separate projects requiring manual switching and content re-pasting between each one.

How does the Notion integration work in this workflow?

Notion connects as an MCP server in Claude Cowork. Once enabled, the agent can query your Notion databases mid-workflow: fetching your published articles list for internal link suggestions, checking your content calendar to identify the next draft to start, or pulling product details for CTA placement. Enable the Notion connector in Cowork’s settings and reference it in your Claude.md instructions.

Key Takeaways

• The problem is orchestration, not tools. Timo’s workflow already had strong individual tools. The friction was managing the handoffs between them. A unified agent eliminates the switching, not the tools.

• Context layer quality determines output quality. Rich, specific context files (About Me, content strategy, target persona) produce more on-brand output than fragmented per-project prompts. This is the most valuable investment in the entire setup.

• Templates and examples do different jobs. Templates give the agent structure (what sections to write). Examples give it tone (how to sound). Both belong in your context layer, but confusing them produces worse results.

• Write your Claude.md in plain English, not code. Dictate your existing workflow out loud, then ask Claude to structure it into an agent file. The more detail you include about your actual process, the better the agent performs.

• The memory layer compounds over time. Every consistent edit you make to the agent’s output becomes a preference it applies automatically. The agent gets better with every article, without you re-explaining preferences.

• The Substack MCP eliminates the last manual step. Direct publishing from Claude to Substack closes the loop on the entire pipeline. From outline trigger to published note, nothing requires leaving Claude.

• The setup cost is front-loaded. Three to five hours of context-layer work, paid once. Every article after runs on that foundation. The ROI calculation is favorable after three articles.

Resources Mentioned

• Claude Cowork: Primary tool demonstrated. The agentic upgrade to Claude Projects, enabling skills, MCP connectors, scheduled tasks, and mobile execution via Dispatch.

• Substack MCP: Custom MCP server for direct Substack publishing from Claude. 11 tools including Notes creation, post management, and subscriber stats. Free for GenAI Unplugged subscribers.

• Notion: Used as article database (for internal linking) and content calendar (for scheduling triggers). Connects via MCP in Claude Cowork.

• Tavily: Research connector for creators with a research phase. Not needed for personal-experience workflows like Timo’s.

• WisprFlow: Mac dictation tool recommended for narrating your workflow into a Claude.md file.

• GPT Image 2.0: Image generation tool Timo uses for thumbnail creation. Post-upgrade, requires 1-2 reprompts instead of 5-7.

• Claude Code: The technical alternative to Claude Cowork. Same 4-layer architecture applies, but agents live in .claude/agents/ folder structure instead of Cowork projects.

Your 15-Minute Challenge

Open a new document and write three things without stopping:

1. Your current AI content workflow, every step, every tool, every copy-paste moment. Don’t organize it yet. Just list it.

2. What you know about your reader that you’ve never written down explicitly.

3. One sentence describing what makes your content different from everyone else in your niche.

That’s the raw material for your context layer. You now have the hardest 20% of the setup done. The rest is connecting it in Claude Cowork and writing the Claude.md file to orchestrate it.

Watch the full session recording to see the live Cowork setup demonstration (starting at 49:30) and the Substack MCP reveal (54:03). The setup takes less time when you can see it happening in real-time.

Success criteria: by the end of 15 minutes, you have a first draft of your About Me file. Everything else follows from that.

Related reading: Build a business-aware AI assistant with Claude Projects

She opened her laptop somewhere over the Pacific, followed a few steps, and had a working content research agent before landing. That one moment sent her on an 18-month journey from one agent to 33.

How to create YouTube Research Agent in Claude Code?

In a recent Substack Live conversation with , we walked through the exact system she uses, including the YouTube Research Agent that replaced her TubeBuddy / VidIQ subscriptions entirely. Kim has been creating content online for 17 years. She is not a developer. And she has never written a line of code for any of her 33 agents.

By the end of this article, you’ll have the full 3-layer architecture to build your own YouTube Research Agent, understand exactly what it produces, know how to control costs from day one, and have a clear first step to get it running this week.

The journey to her 33 AI agents team was made possible through my PubflowOS AI Agents foundational series. It gives you the exact Claude code agent building blueprint that you can replicate to build your own.

Kim Doyal

Build AI-powered solutions without a coding degree. After 17+ years online, I'm showing you how to use AI to save time, scale smarter, and create tools you didn't think you could. Get 'the SPARK' newsletter every Thursday + tutorials and guides in between

Why Are Content Creators Still Paying for YouTube Research Tools?

Most solo content creators running a YouTube channel pay for at least one research tool. TubeBuddy runs $15-35 per month. VidIQ has paid tiers. SEM Rush costs $150 per month or more, and it caps the number of articles or videos it will refresh each month (I was paying $150 per month for a blog refresh service that would only update 12 articles, capped).

The problem is not just the cost. It is that these tools give you generic intelligence.

• They do not know your niche.

• They do not know your audience the way you want.

• They do not know your competitors by name the way you want.

• They do not know your content history or your channel’s positioning.

They give you the same data they give everyone else in your space, which means they give you the same opportunities they give your competitors.

In comparison

A Claude Code YouTube Research Agent costs nothing per month beyond your Claude subscription.

• It pulls real YouTube data through the free YouTube Data API (10,000 units per day, enough for 100 searches daily).

• It knows your specific niche, your named competitors, your content strategy, and your channel’s positioning because you told it once and it remembers.

That’s the shift worth paying attention to here.

Share

What Is a Claude Code Agent, Exactly?

A Claude Code agent is a plain-English instruction file, stored in a folder called .claude inside your project. It tells Claude what to do, which tools to use, which external APIs to call, what context files to load before starting, and what cost limits to respect.

As Kim Doyal put it at 2:30, cutting straight to what confused her friends: “I think of agents as a set of instructions.” That’s it. Not a program. Not a script. Not a piece of software you maintain.

A markdown or YAML file that says: here’s your job, here are your tools, here’s what you know about my business, here’s how many API calls you’re allowed to make.

In the world run by Claude today, you no longer need n8n, AutoGPT, or any agent framework to build working research agents. The architecture lives in three plain-text files.

If you’re new to Claude Code, this primer on building your first Claude Code agent covers the fundamentals before you add the YouTube API layer.

Learn with me: the Claude Code cohort

I run the Claude Code cohort on Maven: six live sessions across six weeks, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Use code GENAI20 for 20% off. Check the Syllabus →

CLAIM 20% OFF

What Does the YouTube Research Agent Actually Produce?

The agent runs in five modes covering the full research workflow. During the live session at 33:30, Dheeraj ran a live demo on the plant-based meals niche. The agent completed the full research cycle in under 60 seconds.

Five modes of YouTube AI Agent

1. --discover: Content gap analysis on a topic. What’s already out there, what’s missing, what angle does not yet exist?

2. --topic [keyword]: Deep research on a specific keyword. Search volume signals, competitor video performance, trending angles.

3. --audit: Channel audit for a specific competitor. Their top performers, their content patterns, their gaps.

4. --shorts: Shorts-specific research. What formats are working, which topics convert to Shorts well.

5. --thumbnails: Thumbnail and title pattern analysis. What title structures are getting the most clicks in your niche.

The live demo output on plant-based meals included: top 5 themes by performance, a list of content gaps, trending signals from the last 90 days, title formats that are outperforming their channel size, and 10 specific video opportunities ranked by a custom signal score.

The content gap discovery

Every beginner guide in the plant-based meals space comes from doctors or single adults. Zero content exists for parents trying to transition their kids. That’s a genuine gap, surfaced in under a minute, by an agent that cost $0 to run.

Custom Signal Score

In the setup guide below, you will see my custom signal scoring formula that correlates views, subscriber count of the posting channel, comment count, and like count into a single multiplier.

This signal score helped me replace the vidIQ subscription that provided guidance on which video to create next.

How Does the 3-Layer Architecture Work?

The 3-layer architecture is the foundation that makes every Claude Code agent reusable and personalizable. Build it once, and adding your 10th or 30th agent takes an afternoon instead of a week.

Kim Doyal built 33 agents from a single foundation. Every one follows the same three layers architecture.

Layer 1: The Data/Integration Layer

This is where external intelligence comes from. For a YouTube Research Agent, the data/integration layer includes:

• YouTube Data API (free, 10,000 units/day): Real video data, not estimates. Actual view counts, subscriber numbers, engagement metrics.

• Perplexity or Tavily (web research MCP): Used for enrichment, competitor analysis, and trend validation after the YouTube API pulls initial data.

• Optional: Notion or Obsidian for connecting your personal content database, video archive, or subscriber research.

The data layer is swappable. If you start with Perplexity and want to switch to Tavily later, you change one line in your MCP configuration file. The rest of the architecture stays the same.

Layer 2: Context Profiles

This is what makes your agent different from a generic research tool. Context profiles are plain-text files stored in your .claude folder:

• business-context.md: Your niche, your audience, your channel positioning, your goals

• content-strategy.md: Your content pillars, your publishing cadence, your existing video archive

• competitor-watch-list.md: Named competitors with their channel links, their strengths, gaps you’ve identified

Every time the agent runs a research session, it loads these files first. It knows who you are before it touches the YouTube API. This is why two people running the same agent on the same keyword get completely different outputs.

Layer 3: The Agent File

A YAML file in your .claude folder. It contains: the agent name, a description Claude uses to decide when to suggest it, a list of tools it’s allowed to use (read, write, search), the model it runs on (Opus for planning, Sonnet for execution), the research modes it supports, and, critically, the cost limits written in plain English.

At 19:57, Dheeraj walked through why this matters: without explicit cost limits in the agent file, Claude will make hundreds of YouTube API calls in a single session. You hit your daily free quota in minutes. Writing the limit once prevents that permanently.

Get PluggedIn

Two non-developers replaced $50/month in YouTube research subscriptions with a 3-file Claude Code agent. No code written.

See what’s included →

Every month you stay with SEM Rush, TubeBuddy, or VidIQ, you’re paying up to $15-50 for data that’s identical to what your competitors are buying. Niche-agnostic. Capped. The $150/month I was paying for a service that refreshed just 12 articles is the cost of not having a personalized agent.

Get PluggedIn to go from paying for generic research data to running a YouTube Research Agent that knows your niche, your competitors by name, and your channel’s positioning. The complete agent file, MCP configuration, context profile templates, and custom signal scoring formula from this session are inside.

Read more

That’s not meant to scare you. It’s the reality of OpenClaw, Claude Dispatch, and Claude Code Channels. I recently talked with Chris from . He’s one of the few people who has done serious security research on AI agents. Together, we went deep on what these new remote access features actually do.

ToxSec - AI and Cybersecurity

Security for a world run by machines that lie.

The timing matters. Claude Dispatch, Claude Channels, and OpenClaw all became mainstream options within a few weeks of each other. Most creators using these tools have no idea what permissions they’re granting or what happens when something goes wrong.

By the end of this post, you’ll understand the real security model behind each tool. You’ll learn the one attack type that affects all three. And you’ll get a practical playbook for using these tools safely.

The tools are genuinely useful. The goal isn’t to stop you from using them. The goal is to make sure that when something goes wrong (and statistically, it will), the damage is contained to something you can recover from. That’s the only security framework that works at solo scale.

What Is Claude Dispatch and What Permissions Are You Granting?

Claude Dispatch is a remote shell, not a messaging app. That distinction matters more than it might seem.

When you pair your phone with your desktop using the QR code flow, your phone gets every permission that Claude Code has on your desktop. That includes browser access, MCP server tools, file system access, and any integrations you’ve connected. The QR code setup is well-built. Traffic runs over TLS through Anthropic’s API. Chris is clear on this at [3:13]: the pipe itself is secure.

The risk lives in what sits on the other end of that pipe.

“You lose your phone with no password, and they can suddenly send all sorts of stuff on your home computer.” - Chris

At [5:41], Dheeraj showed this live during the session. He sent a command from his phone while on camera. Claude opened Chrome on his desktop and reported back on his LinkedIn impressions. The capability is real, it works, and it’s genuinely useful.

Here’s the security reframe: your phone’s passcode is now your workstation’s security perimeter. If your phone has a weak password - or no password - and you lose it, that person can send commands to your home computer.

What this means practically:

• Strong Face ID or passcode is no longer optional if Dispatch is active

• Every integration you’ve connected to Claude Code is now reachable from your phone

• There is no per-task permission separation. You cannot say “only allow GitHub access from mobile.” It’s all or nothing.

What Is Prompt Injection and Why Does It Affect Every AI Agent?

Prompt injection is the core vulnerability of every AI agent with real-world capabilities. No exception.

Here’s how Prompt Injection works:

LLMs have no built-in separation between “instructions from the user” and “data being processed.” It’s all one context window. When your agent reads an email to summarize it, the text inside that email can contain instructions. The model treats those instructions as real commands.

As Chris explains at [9:21], the attacker doesn’t need to be technical. Here’s one example: an attacker sends you an email with white text at font size 0.1. You see a normal email. The model reads the HTML and treats the hidden text as real instructions.

“Whatever folder you have it operating in, those files are under attack.” - Chris

The instructions might say: summarize this email, then forward these files to attacker@domain.com. Or: add this contact to your address book. Or: post this message to LinkedIn. The agent follows them because it can’t tell them apart from your real instructions.

There is no alarm. No notification. Most users would never know it happened.

The more integrations your agent has, the more damage a successful injection can cause. An agent with read-only file access is much safer than one connected to Gmail, LinkedIn, and a browser. Each integration is a new attack surface and makes the damage worse.

This is why Chris’s core framework is blast radius reduction, not injection prevention. You will eventually get injected. Build for that assumption.

Subscribe to ToxSec

Learn with me: the Claude Code cohort

I run the Claude Code cohort on Maven: six live sessions across six weeks, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Use code GENAI20 for 20% off. Check the Syllabus →

CLAIM 20% OFF

Claude Channels vs. Dispatch: Understanding the Blast Radius Difference

Dispatch and Channels are both Anthropic products, but they work at very different levels of system access. Understanding this gap shapes every decision about how to use them.

At [15:59], Dheeraj showed the Channels setup screen live. Something unusual is right there in the interface. Anthropic itself shows a prompt injection warning before you turn it on. That’s a clear signal from the company building the product.

Here’s the core difference:

Claude Dispatch

• Operates at the app level (somewhat sandboxed)

• Inherits Claude Code permissions

• Active only when you’re actively using it

• Blast radius: your Claude Code session

Claude Channels (Telegram, Discord, iMessage)

• Operates at full system level

• Exposes bash and your full working directory

• Runs whenever a message triggers it

• Blast radius: your entire computer

Chris’s explanation at [19:00] makes the difference concrete. If you visit a hacked website while Channels is active, that site’s content can inject commands into your whole system. Not just your Claude Code session. The whole machine.

Many in the community see Channels as Anthropic’s answer to OpenClaw’s growing adoption. They shipped it with a built-in injection warning. That tells you they know the risk.

The practical takeaway: if you’re using Channels for Telegram, treat every message your agent reads as a potential attack. That doesn’t mean avoid it. It means the files and integrations that agent can reach should be ones you’d be fine losing.

Share

This post gives you the what. Building the actual five-layer isolation setup (hardware choice, VPN config, permission boundaries, burner accounts, hardening checklist) still takes most solo creators 4 to 6 hours of piecing it together from GitHub issues and scattered docs.

PluggedIn members get the exact config checklist and setup template Dheeraj runs, so the assembly step is already done for you.

OpenClaw: Where Self-Hosting Adds Power and Removes Protection

OpenClaw runs 24 hours a day, 7 days a week. That’s its core value and its core security risk.

At [26:05], Chris is direct:

“OpenClaw is a 24/7 prompt injection engine. It is just waiting to get prompt injected.” - Chris

This isn’t a knock on OpenClaw. It’s just the reality of any always-on, internet-connected agent.

Where OpenClaw gives you more:

• Data boundary control. Sensitive information never leaves your setup.

• Local model routing. You can run sensitive requests through local models. This cuts token costs and keeps data out of Anthropic’s systems.

• No platform guardrails limiting what it can do.

Where OpenClaw removes protection:

Anthropic’s managed setup has two protection layers that go away when you self-host. The first is model-level guardrails. These are Claude’s built-in refusals, trained through reinforcement learning (RLHF).

The second is platform-level guardrails. These are separate controls that can block dangerous commands before the model sees them. Some systems even review the model’s outputs before they run.

When you run OpenClaw, you keep layer one (the model) but lose layer two (the platform controls). As Chris explains at [31:20], most people do worse with more control, not better.

There’s one more serious issue. OpenClaw can read and modify its own configuration file.

A successful injection doesn’t just run a command. It can rewrite OpenClaw’s own permissions and give itself more access. Chris explains at [24:07] that this is why hardware isolation isn’t optional for OpenClaw.

Chris’s own setup: a dedicated Linux mini PC used only for AI agent work. Nothing sensitive on the machine. Fully isolated from his main computer.

The Practical Security Playbook for Solo Creators

At [44:22], Chris shares the most useful part of the session. The framework is assume breach. Design for the moment when an injection happens, not to prevent it.

“Assume it’s going to be prompt injected. Chances are it will be. So if you build your product under the assumption that this thing will get hit, then you just limit what’s the damage that could be done.” - Chris

Here’s the tiered approach:

Tier 1: Dedicated hardware or VPS (strongest isolation)

• Spin up an AWS EC2 instance or similar VPS. Relatively cheap for light agent workloads.

• Or use dedicated hardware: Raspberry Pi (~$100-150 for a complete kit), Mac Mini, or Linux mini PC.

• The agent lives on that machine. Nothing else does.

Tier 2: Virtual machine (minimum viable isolation)

• If you can’t justify dedicated hardware, a VM on your existing computer limits the blast radius at almost no cost.

• A successful injection stays inside the VM.

For integrations, regardless of tier:

• Give the agent its own email account. Not your personal Gmail. A burner address used only for the agent’s tasks. An attacker who injects through that inbox can only reach what that inbox touches.

• Limit integrations. Each connected tool is a new injection surface and a new blast radius.

• GitHub for code pushes is lower risk. Your main email with years of personal data is not.

“Everything on that machine is something you would be comfortable giving to a stranger, because effectively you’re going to give everything on that machine to a stranger.” - Chris

This mental model is useful. Before you connect any integration, ask: would I be fine if a stranger had access to this? If not, don’t add it.

After this setup, an injection through your agent’s email only reaches one burner inbox. Not your main accounts. Not your files. Not your main computer.

Leave a comment

Structured Security Comparison: Claude Dispatch vs. Channels vs. OpenClaw

Claude Dispatch

• Access type: Remote shell via phone

• Always on: No (active sessions only)

• System access: App-level (sandboxed)

• Managed guardrails: Yes (Anthropic platform + model)

• Injection surface: Phone messages, browsed content

• Best for: Occasional remote commands, lower-sensitivity tasks

Claude Channels (Telegram, Discord, iMessage)

• Access type: Messaging platform integration

• Always on: Triggers on message receipt

• System access: Full system level (bash, working directory)

• Managed guardrails: Yes (with explicit Anthropic warning)

• Injection surface: Every message processed

• Best for: Power users with isolated agent setup

OpenClaw

• Access type: Self-hosted web UI + heartbeat

• Always on: Yes (24/7 process)

• System access: Full system on host machine

• Managed guardrails: Model only (no platform layer)

• Injection surface: Web content, docs, all ingress data

• Best for: Technical users on isolated dedicated hardware

Share GenAI Unplugged

Get PluggedIn

You know the blast radius problem. Now get the five-layer setup that actually contains it.

See what's included →

Without the config template, those 4 to 6 hours of setup work are still ahead of you, or you keep running these tools exposed until something forces your hand.

Get PluggedIn to go from understanding the risks but running an unhardened setup to having the five-layer isolation configured and working in one afternoon

Frequently Asked Questions

Is Claude Dispatch safe to use?

The transmission layer is secure: TLS encryption, QR code authentication, Anthropic’s API. The risk is that your phone gets all Claude Code permissions from your desktop. Losing an unlocked phone means someone can control your home computer. Use strong phone authentication and check what permissions Claude Code has before you turn on Dispatch.

What is prompt injection in AI agents?

Prompt injection is when an attacker hides malicious instructions inside data that an AI agent reads. LLMs have no built-in separation between user instructions and data being processed. A common method: hidden white text (font size 0.1) in an email. The model reads it as real instructions, but you can’t see it.

Can I limit what Claude Dispatch can access from my phone?

Not currently. There is no per-task or per-device permission separation in Dispatch. Whatever permissions Claude Code has on your desktop are fully passed to Dispatch on your phone. The only way to limit scope is to restrict permissions in your base Claude Code setup before turning on Dispatch.

Should I use OpenClaw or Claude’s managed Dispatch and Channels?

For non-technical users, Anthropic’s managed tools are generally safer because you keep both model-level and platform-level guardrails. OpenClaw offers data privacy and more control but you lose the platform protection layer. Technical users who understand blast radius and can run OpenClaw on isolated hardware can use it safely. OpenClaw’s default setup shipped insecurely; don’t run it without hardening it first.

What is the safest setup for running AI agents remotely?

Tier 1: Dedicated VPS (AWS EC2) or dedicated hardware (Raspberry Pi, Mac Mini), fully isolated from your main computer. Tier 2: A virtual machine on your existing hardware. In both cases, use separate burner accounts for each service. Never use your main email or primary accounts. Adopt an assume-breach mindset and limit integrations to only what you need.

Is it safe to give Claude access to my Gmail for email summarization?

Not without careful setup. Anyone who emails you could inject commands into your agent using hidden text. If you want email summarization, create a separate email account for it. Don’t connect your main inbox. This limits the damage if an injection happens.

Key Takeaways

• Phone security equals workstation security. When Dispatch is active, a lost unlocked phone is a full desktop breach. Strong Face ID or passcode is non-negotiable.

• Prompt injection affects every agent with real-world capabilities. There is no patch. The fix is blast radius reduction, not prevention.

• Channels has a larger blast radius than Dispatch. Full system-level access vs. app-level. Anthropic ships it with a built-in injection warning for a reason.

• OpenClaw is powerful but requires isolated hardware. It runs 24/7, can modify its own config, and you lose Anthropic’s platform-level guardrails. Don’t run it on your main machine.

• Assume breach. Design every agent setup under the assumption that an injection will eventually succeed. The question is what the attacker gets when it does.

• Integration count equals attack surface. Each connected tool is a new injection vector. Limit integrations to only what you actively use.

• Separate accounts for agent integrations. Burner email, not your main Gmail. Agent GitHub, not your personal account. Limit what an injection can reach.

• Used correctly with isolated hardware, each of these tools genuinely extends your capability as a solo creator running AI systems at scale.

Subscribe now

Resources Mentioned

• Claude Dispatch - Anthropic’s remote shell feature enabling phone-to-desktop Claude Code control via QR code authentication

• Claude Channels - Anthropic’s integration layer for external messaging platforms (Telegram, Discord, iMessage) with full system-level access

• OpenClaw - Self-hosted Claude agent with web UI, always-on heartbeat process, configurable model backends

• OpenClaw Security Checklist - Chris’s published hardening guide for OpenClaw; referenced by Dheeraj as the basis for his five-layer setup

• Tailscale VPN - Dheeraj’s VPN layer in his OpenClaw setup, used alongside four other security configurations

• IronClaw - Security-focused OpenClaw alternative, built in Rust, described by Chris as “supposed to be really good”

• AWS EC2 - Recommended cloud VPS option for isolating agents from main computers

• Raspberry Pi - ~$200 budget hardware option for dedicated agent hosting

• SOC 2 Type 2 / ISO 27001 - Anthropic security certifications cited as evidence of managed platform’s security posture vs. self-hosted alternatives

Leave a comment

Your 15-Minute Challenge

Open your Claude Code settings and audit every integration you’ve connected. For each one, ask: if this integration was compromised, what could an attacker reach?

Make three decisions:

1. Keep integrations where the blast radius is acceptable (a dedicated GitHub account, a burner email address)

2. Disconnect integrations that expose your primary accounts or sensitive files

3. Create a burner email account for Claude agent tasks, and swap out your main email if it’s currently connected

This audit takes 15 minutes. It’s the best security action you can take before enabling Dispatch or Channels.

For the full conversation - including the live Dispatch demo and Chris’s breakdown of Anthropic’s dual guardrail system - watch the full session recording.

Next up: The exact five-layer OpenClaw setup Dheeraj runs, including hardware, VPN, permissions, and the one config change that matters most.

That’s the honest starting point. Not a hypothetical. Not an edge case. For me, solo creator and host of GenAI Unplugged, thumbnail creation was always the last thing standing between a finished video and a published one.

The video was edited. The description was written. But the thumbnail wasn’t ready. So I’d open ChatGPT or Claude for creative direction, then Canva for execution, then spend 30+ minutes hoping the result looked consistent with the rest of his feed.

When Google Gemini Nano Banana Pro launched, that changed. The result is a 5-stage n8n pipeline that goes from video transcript to finished thumbnail in under 5 minutes, costs roughly $0.18 per image, and runs reliably enough that I no longer needs a designer.

In a recent live conversation with from Nine to Thrive, I walked through the full pipeline live, including shortcuts for solo creators who want fewer tools and a code-first alternative for those who prefer to stay local.

9-to-Thrive

Offers that sell. Mornings that don't. Every week, I help creators figure out what to sell, how to price it, and build revenue that shows up whether they're working or not. The exact systems I used to build 6 income streams and $20K/month in 3 hours/day.

By Carrie Loranger

By the end of this article, you’ll understand exactly how the pipeline works, what each stage does, how much it costs, and how to decide which version makes sense for your setup.

The pipeline is 90-95% reliable. At worst, you regenerate the thumbnail with one click. It takes two minutes and produces a completely fresh concept. The moment Gemini Nano Banana Pro launched, things changed. I just don’t think about thumbnails much anymore.

Why Is Thumbnail Creation the Last Bottleneck Worth Automating?

For solo YouTube creators publishing 4-8 videos per month, thumbnails eat 2-5 hours every month a time sink hiding inside a task that feels like it should be quick. The real problem is not the time alone.

It’s that every thumbnail demands both creative direction and visual execution, and manual workflows force you to do both under deadline pressure.

At 5:49 in the session, Dheeraj describes the old reality:

“I would spend about 30 to 40 minutes just figuring out what thumbnail is needed. And thumbnail is not ready by the time I have to publish. So that’s where I started thinking.” - Dheeraj

The pipeline described here collapses that two-part problem into a single automated pass. AI handles the creative direction by reading the full video transcript. A separate AI pass handles the visual execution using your avatar reference image and brand rules. You review the result. If it’s good (and it usually is), you’re done.

If it’s not, you regenerate. Two minutes. Fresh concept. No design tools required.

How Does the 5-Stage Thumbnail Pipeline Work?

The full architecture runs from raw video to finished thumbnail in five stages. Each stage handles one piece of the problem so you don’t have to. Here’s how n8n orchestrates the workflow from start to finish. If you’re new to n8n, our free 47-lesson course covers everything from installation to advanced workflows.

Stage 1: Image Queue

A Google Sheet acts as the control panel. Each row contains a YouTube video ID and a status column with one of three values: scheduled, new upload, or thumbnail created. The workflow filters for new upload rows and uses a Limit node to process one video at a time.

Stage 2: Transcript Fetch

This is where most of the setup complexity lives, and where the solo shortcut matters most. The default path: upload your video to YouTube as unlisted (not draft, not published), wait 30-40 minutes for auto-generated captions, then fetch the transcript via the YouTube Transcriptor API on RapidAPI (free tier: 50-100 requests per month). A code node then runs ASR cleanup, correcting misspelled tool names and domain-specific vocabulary that auto-captioning gets wrong.

Stage 3: AI SEO Brain

The cleaned transcript feeds into a Gemini 3 Pro model (or Claude) with a detailed system prompt covering brand voice, headline rules, and thumbnail context format. The system prompt structure follows the same principles we covered in prompt engineering for automation. In one pass, this stage produces: a video title, description, hashtags, and three ranked thumbnail concepts with CTR rationale. At 27:00, Dheeraj calls this the creative core:

“This is where the creativity is happening. You are leveraging AI to get that creative angle for your thumbnail.” - Dheeraj

A parallel branch from this same stage can feed the AI-generated title, description, and tags directly to a YouTube Update node, automatically updating the video’s metadata before it goes public.

Stage 4: Director Prompt

The thumbnail-relevant fields pass through a second AI stage: the director prompt. This stage selects a pose from a predefined library (12-13 options: shocked, pointing, confident, teaching, thumbs up, stop) and writes the final image generation prompt using a 4-layer composition system: abstract background, subject at 80% right, floating 3D icons, text overlay. The identity lock technique (more on this below) runs here.

Stage 5: Image Generation and Upload

Your avatar reference image and any background references are converted to base64. These feed into the Gemini Imagen API call alongside the composed prompt, at 1K resolution in 16:9 ratio. The output is a base64-encoded image that gets decoded and uploaded to a designated Google Drive folder. If the API returns no image, an if-node catches it and loops back for up to three retries before stopping.

Real cost data: $0.15-$0.18 per thumbnail at 1K resolution. Gemini’s published rate is $0.10, but real-world API bills run slightly higher. At 8 videos per month, that’s under $1.50/month in image generation costs.

Learn with me: the Claude Code cohort

I run the Claude Code cohort on Maven: six live sessions across six weeks, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Use code GENAI20 for 20% off. Check the Syllabus →

CLAIM 20% OFF

Get PluggedIn

Stop spending 30-40 minutes per thumbnail and ship with a 5-minute automated pipeline instead.

See what’s included →

Next video publish: another 30-40 minutes staring at Canva, hoping it looks consistent.

Get PluggedIn to go from manually wrestling with thumbnails before every upload to running a 5-stage n8n pipeline that delivers a finished image in under 5 minutes

What If You Want Fewer Tools? The Solo Creator Shortcut

At 23:22, Carrie surfaces something most tutorials skip. Solo creators don’t always want the full API setup - they want the result with fewer moving parts. The shortcut eliminates three tools from the pipeline while keeping the AI stages intact.

“As a solo creator, I like to limit the number of tools and logins that I have.” - Carrie

The full pipeline requires a RapidAPI account, a YouTube Data API key, and the unlisted video workflow. That’s three new credentials for someone who just wants their thumbnails automated.

The shortcut: add a transcript column to your Google Sheet and paste your transcript text directly into the cell. The AI stages downstream don’t care where the transcript came from. You can also point the pipeline to a Google Doc instead of a YouTube API call.

This eliminates: the unlisted video requirement, the RapidAPI account, the YouTube transcript API call, and the ASR cleanup node.

The trade-off is one small manual step: pasting the transcript. For creators who already have a transcript from their recording tool or who use Substack Live with auto-captions, this is actually less work, not more.

Leave a comment

How Do You Get AI-Generated Thumbnails to Actually Look Like You?

The most common failure mode in AI thumbnail generation is identity accuracy. The identity lock technique - repeating your avatar description at both the top and bottom of a long prompt - cuts miss rates from 3-in-10 down to roughly 1-in-10.

Combined with a clean reference image, this makes character consistency reliable enough for production use. The AI produces a great composition, but the person in the thumbnail doesn’t look quite right.

At 37:00, Dheeraj shares the technique behind the identity lock.

The technique is simple. In your image generation prompt, describe your avatar in detail at the very top of the prompt. Write all of your composition instructions in the middle. Then repeat the avatar description again at the very bottom of the prompt.

The repetition anchors the model’s attention to your identity at both the start and end of its context window. Combined with providing a clean reference image (upper body only, transparent or solid background, no distracting elements), this single change makes a measurable difference.

“I only provide the upper part of my image so the book is not there. Accuracy started improving when I only provided my image without the destructive elements.” - Dheeraj (at 43:47)

Your reference image matters as much as your prompt. If you’re getting inconsistent results, improve the reference image before adjusting the prompt.

If you publish 4-8 videos a month, manual thumbnails are costing you 2-5 hours every month and that's before you factor in the deadline pressure that kills creative quality.

PluggedIn has the exact n8n workflow template, prompt library, and pose reference system from this pipeline so you can skip the build-from-scratch part entirely.

How Do You Debug and Fine-Tune the Thumbnail Prompts?

When a prompt isn’t producing what you want, the fastest path to fixing it is multimodal debugging - not blind trial and error. Feed both the prompt output and the generated image to Claude or ChatGPT and ask it to identify the gap between intent and result. This approach is similar to how we use AI to audit content quality in our pipeline. At 32:53, Dheeraj describes his actual debugging workflow:

“I would dump the entire output of all the nodes into Claude, put the image in Claude as well and ask: what do you see? What is wrong? What am I missing? And it would help me fine-tune the entire prompt over time.” - Dheeraj

This works because Claude can analyze both the text output (thumbnail context, prompt) and the visual result (generated image) simultaneously. It can identify gaps between what the prompt asked for and what the model produced, and suggest specific edits to close that gap.

This same approach applies to tuning the AI SEO Brain prompt for your niche. Carrie suggests pointing Claude at your existing brand voice documents and asking it to rewrite the system prompt to match your style. You paste the result back into the n8n node. No manual rewriting required.

Join Carrier at Nine to Thrive

Can You Build This Without n8n Using Claude Code Instead?

Yes - for creators who prefer working in the terminal or who already use Claude Code, a code-first version produces the same result. Claude Code skills combine a markdown prompt file with a Python script, giving you the full pipeline without a visual builder. This approach works well if you’ve already set up Claude Code for your content workflow.

At 54:27, Dheeraj demonstrates the yt-thumbnail skill: a markdown skill file that defines brand colors, poses, and the system prompt, backed by a Python file (generate_thumbnail.py) that handles the Gemini API call. The character reference image is stored in the global Claude directory, making it accessible to all skills without re-uploading.

To trigger it: describe what you want in plain English and provide the path to your transcript file. Claude Code handles the rest.

The accuracy in the live demo was slightly better than the n8n version for that particular session, likely because the skill’s identity description was more dialed in.

Which approach makes sense for you:

n8n

• Best for: Visual workflow builders, non-technical creators, those who want to share workflows with a team

• Setup: 10-15 minutes (local) or 5 minutes (cloud at $24/month)

• Runs without you: Yes, on a schedule or trigger

• Cost: n8n subscription + $0.15-0.18/image

Claude Code yt-thumbnail skill

• Best for: Technical creators, those already using Claude Code, local-first setups

• Setup: Install Claude Code + configure skill + store reference image

• Runs without you: No (requires a prompt to trigger)

• Cost: Claude Max subscription + $0.15-0.18/image

Frequently Aaked Questions

How long does it take to set up the n8n thumbnail automation workflow?

The n8n installation takes 10-15 minutes for the free community edition or about 5 minutes for the cloud version ($24/month). The main setup time is configuring credentials: Google Sheets, YouTube Data API, RapidAPI, Gemini API, and Google Drive. Credentials set once work across all future workflows. If you use the solo shortcut (manual transcript paste), you skip the YouTube API and RapidAPI credentials entirely.

How much does it cost to auto-generate YouTube thumbnails with AI?

Approximately $0.15-0.18 per thumbnail using Google Gemini Image Pro at 1K resolution in 16:9 format. Gemini’s published rate is $0.10, but real-world API bills run slightly higher. At 8 videos per month, expect roughly $1.20-$1.44 in image generation costs per month.

Why does the video need to be unlisted on YouTube instead of a draft?

The YouTube Transcriptor API (via RapidAPI) can only access transcripts from publicly reachable URLs. Draft videos are private to your account and inaccessible to external APIs. Unlisted videos are accessible via direct URL without notifying subscribers, which is why unlisted is the right status for this workflow.

What if I don’t want to use the YouTube API for transcripts?

Add a transcript column to your Google Sheet and paste your transcript text there. The pipeline’s AI stages use whatever transcript text they receive and don’t care about the source. You can also use a Google Docs node pointing to a document where you paste or import your transcript. Either option eliminates the RapidAPI and YouTube API requirements.

How accurate is the AI at generating thumbnails that look like me?

About 90-95% of thumbnails come out usable without changes. The most common issue is identity accuracy, where the avatar doesn’t quite look like you. Using the identity lock technique (repeating your avatar description at the top and bottom of the prompt) and providing a clean reference image (upper body only, transparent or solid background) brings the miss rate down to roughly 1-in-10. Regenerating takes 2 minutes and produces a fresh concept.

Can this pipeline also update my YouTube video title, description, and tags?

Yes. Add a YouTube Update node as a parallel branch from the AI SEO Brain stage (Stage 3). The same AI pass that generates thumbnail concepts also produces a video title, description, and hashtags. Feed those directly to the YouTube node to update the unlisted video before it goes live.

Can this pipeline work for Substack cover images too?

Yes. Dheeraj runs two separate versions: one tuned for YouTube thumbnails and one for Substack covers. The core pipeline is identical. The difference is in the brand rules, composition instructions, and aspect ratio in the director prompt stage.

Key Takeaways

• The 5-stage pipeline compresses thumbnail creation from 30-40 minutes to under 5 minutes at a cost of $0.15-0.18 per image, including creative direction and visual execution.

• The unlisted YouTube video status is a prerequisite for transcript API access. Draft mode blocks external API calls. Unlisted mode allows them without notifying subscribers.

• The AI SEO Brain stage is dual-purpose. One AI pass generates thumbnail concepts AND video metadata (title, description, tags), both of which can be used automatically downstream.

• The identity lock technique cuts identity misses by roughly two-thirds. Repeat your avatar description at the top and bottom of the image generation prompt, and provide a clean upper-body reference image.

• The solo shortcut (manual transcript paste) removes three tools from the setup. If you want the benefits without the YouTube API complexity, add a transcript column to your Google Sheet.

• Debugging with Claude saves hours of trial and error. Dump all node outputs plus the generated image into Claude and ask what’s wrong. It identifies gaps between prompt intent and image output more accurately than manual review.

• n8n and Claude Code both produce the same result via different paths. n8n is visual and shareable. Claude Code is local and integrates with your existing development workflow.

Resources Mentioned

• n8n - The visual automation platform used to build the pipeline. Free community edition (local install) or $24/month cloud version.

• Google Sheets - Queue management for video IDs and status tracking.

• YouTube Transcriptor API via RapidAPI - Fetches auto-generated transcripts from unlisted YouTube videos. Free tier: 50-100 requests/month.

• Google Gemini 3 Pro - AI model used for the SEO Brain stage (transcript analysis, thumbnail concepts, video metadata).

• Google Gemini Image Pro - Image generation model for the final thumbnail. $0.10 published rate, $0.15-0.18 real-world cost per image at 1K resolution.

• Google Drive - Final destination for generated thumbnail files.

• Claude - Recommended for debugging pipeline outputs, rewriting brand prompts, and as an alternative AI model in the SEO Brain stage.

• Claude Code - Code-first alternative to the n8n pipeline, demonstrated via the yt-thumbnail skill.

• Ideogram - Mentioned as an alternative image generation model that accepts reference images.

What Can Go Wrong? Common Issues and Fixes

AI Thumbnail doesn’t look like you

Start with a better reference image. Crop to upper body only, use a transparent or solid-color background, and remove any props (books, logos, accessories). Then apply the identity lock technique: repeat your avatar description at the top and bottom of the prompt. This alone fixes most identity misses.

Gemini API returns no image (empty response)

The retry loop handles this automatically (up to 3 retries). If it keeps failing, check your Gemini API quota in the Google Cloud console. Free-tier limits can cause silent failures during peak hours.

ASR cleanup misses tool names

Add new correction entries to the Code node. The format is a simple find-and-replace array. Common additions: your brand name, niche-specific tools, and co-host names that speech recognition mangles.

n8n workflow errors on first run

Most first-run errors are credential issues. Verify each Google credential (Sheets, Drive, YouTube) is authorized in your n8n instance. Credentials set once persist across all workflows. Check the n8n credentials documentation for step-by-step setup.

Prompt produces wrong visual style

Dump the full node output plus the generated image into Claude and ask: “What’s the gap between my prompt and this result?” Claude can identify specific prompt instructions the model ignored and suggest targeted edits. Two rounds of this typically dial in the style.

Share GenAI Unplugged

Your 15-Minute Challenge

Pick the version that matches your current setup.

If you use n8n (or want to):

Create a Google Sheet with three columns: video_id, status, and transcript. Add one row with a video you’ve already published, set the status to new upload, and paste in its transcript. That’s your queue. You now have the foundation to plug in every other stage of the pipeline.

If you use Claude Code:

Open a new markdown file, name it yt-thumbnail.md, and write three things: your avatar description (physical appearance, style, expression), your brand colors, and one example thumbnail concept that worked well for you. That’s the start of your skill file.

Both versions start with the same thing: defining what you look like and what your brand sounds like. Do that first. The pipeline is ready when you are.

Get n8n Workflow (50% OFF)

• The 17-year content journey - From a Himalayan travel blog (Devil on Wheels, 562 articles) helping local porters and homestays, to a YouTube channel, to building AI automation systems that power everything today.

• The “Automate, Batch, Compress” framework - Dheeraj’s personal productivity mantra. Automate repetitive tasks with AI and n8n, batch-create content in series (not one article at a time), and compress work into short focused sessions because Claude Code retains all the context.

• How SubflowAI was born - Started as a personal n8n workflow for Substack Notes scheduling. When the workflow broke due to Substack API changes, Claude Code suggested building a Chrome extension instead. Built the first version in a single weekend. Now has 20 beta users.

• The 3-tool Content Operating System - Notion as the persistent database, Claude Code as the brain that builds and orchestrates everything, and n8n for external integrations and distribution. Everything built with Claude Code in weekend sessions.

• Practical advice for beginners - Don’t start by building agents. Start with a Claude Project, upload your brand voice, and write one article per chat. The need for automation will come naturally when you find yourself repeating the same steps.

Learn with me: the Claude Code cohort

I run the Claude Code cohort on Maven: six live sessions across six weeks, building a real AI system with me. If you would rather build it alongside me than read about it, this is the room.

Each cohort is a small, hands-on group, so seats are limited.

Use code GENAI20 for 20% off. Check the Syllabus →

CLAIM 20% OFF

Key Moments

• 00:00

  • Topic: Welcome & intro - Vinayak congratulates Dheeraj on 300 subscribers

• 00:57

  • Topic: Dheeraj’s intro - AI systems, automation, 17-year content creator journey

• 02:49

  • Topic: The origin story - Travel blog “Devil on Wheels” for Indian Himalayas, helping local porters & homestays get discovered

• 04:53

  • Topic: Transition to YouTube - Overcoming camera shyness, took 1.5 years to record first video

• 06:57

  • Topic: First AI use case - RAG chatbot to handle 60-70 daily YouTube comments

• 09:52

  • Topic: The batching discovery - Planning content around pillars, batch creating, reducing context switching

• 10:39

  • Topic: “I automate, I batch, and I compress” - Dheeraj’s personal mantra for productivity

• 11:03

  • Topic: Claude Code as the go-to tool - Replacing a lot of tools with one AI coding assistant

• 15:07

  • Topic: Shifting focus to GenAI Unplugged - Why the AI/automation channel started

• 15:48

  • Topic: Gap in n8n education - Most videos showed demos, not structured zero-to-hero learning

• 17:22

  • Topic: Teaching as learning - “Whatever I’m learning, I’m going to teach”

• 19:50

  • Topic: Starting with the pain - YouTube thumbnail creation as first automation project

• 22:28

  • Topic: Building systems with Claude Code in a weekend - Contrast with travel blog days where no systems existed

• 22:41

  • Topic: LinkedIn FOMO & discovering Substack - Draining energy on LinkedIn, finding Substack’s curated community

• 25:40

  • Topic: Content pipeline shift - From YouTube-first repurposing to Substack-native content

• 27:00

  • Topic: Build in Public approach - Solving real problems (neighbor’s job hunting, restaurant invoices) then sharing the journey

• 29:03

  • Topic: How SubflowAI was born - n8n pipeline broke, Claude Code suggested a Chrome extension, built in a weekend

• 33:42

  • Topic: SubflowAI features - Batch scheduling, content repurposing, scheduling strategies

• 35:48

  • Topic: Migrating n8n from n8n Cloud to free Oracle VM - Claude Code scripted the entire migration, saving $25/month

• 37:34

  • Topic: Oracle free tier hack - Claude Code automated a retry script (41 attempts!) to grab a 24GB RAM machine

• 39:17

  • Topic: Claude Code Max plan ROI - Saves $60-70/month in hosting alone, built 3 products, replaces $165/month Ahrefs

• 42:03

  • Topic: Self-hosted n8n vs VPS vs local - Detailed comparison for the audience

• 44:30

  • Topic: Self-hosting everything on Oracle - Website, n8n, Gotenberg PDF service, all managed by Claude Code

• 45:55

  • Topic: Markdown-to-branded-PDF pipeline - /pdf Claude skill for product documentation

• 46:07

  • Topic: File sync utility - Built in 2 hours with Claude Code to replace $20/month backup tool

• 48:42

  • Topic: Non-technical friend builds first app with Claude Code - Restaurant invoice automation

• 50:29

  • Topic: Advice for beginners - Map your process, rate pain areas by ROI, target the biggest friction first

• 53:19

  • Topic: Simplest starting point - Claude Project with brand voice uploaded, one chat per article

• 57:24

  • Topic: Natural progression to Claude Code - When you catch yourself repeating interactions in Claude Projects

• 59:30

  • Topic: The 3-tool content OS stack - Notion (database) + Claude Code (brain) + n8n (external integrations)

• 1:01:07

  • Topic: Avoiding tool FOMO - Why Dheeraj stays focused on shipping, not chasing every new AI tool

• 1:04:30

  • Topic: Closing - Subscribe to GenAI Unplugged YouTube & Substack

Dheeraj cut $225/month in SaaS costs using Claude Code and a free Oracle VM. Reading about the Oracle retry script and n8n migration gives you the concept. Replicating it without the actual scripts and configs costs you weeks of trial and error, or the $225 keeps leaving your account.

The exact retry script, migration configs, and deployment checklists Dheeraj used are packaged inside PluggedIn so you skip the rebuilding.

Top Takeaways

1. Start with the pain, not the tool - Rate your pain areas on a scale of

2. Automate the highest-ROI friction first. For Dheeraj, it was YouTube thumbnails, then Substack Notes scheduling.

3. Systems beat willpower - The travel blog failed to scale because there were no systems, only theoretical batching. With AI, you can build real systems in a weekend that handle the repetitive parts.

4. Claude Code pays for itself - Dheeraj migrated n8n Cloud ($25/mo) and DigitalOcean ($35/mo) to a free Oracle VM, replaced Ahrefs ($165/mo), and built 3 products - all with Claude Code. The Max subscription ROI is clear.

5. You don’t need to be technical - A restaurant owner built his first invoice automation app by just explaining the problem in plain English to Claude Code. If you can articulate the problem, AI can build the solution.

6. Avoid FOMO, protect your momentum - Dheeraj consciously avoids chasing every new AI tool. He doesn’t publish tool reviews. He stays focused on building and shipping what he needs.

7. The natural progression path for Dheeraj - ChatGPT -> Claude Desktop with Projects -> Claude Code. Don’t force the jump. The need for automation surfaces naturally when you catch yourself repeating the same chat interactions.

8. Build in public, learn by teaching - Every system Dheeraj builds for himself or friends becomes content. The building IS the content strategy.

People / Resources Mentioned

• WhisperFlow - Voice dictation tool Dheeraj uses instead of typing (Use my link, you get one month FREE to try, I get one month free too :))

• Claude Code - AI coding assistant (Anthropic) - core of Dheeraj’s content OS

• n8n - Open-source workflow automation platform

• Notion - Database layer for the content operating system

• Oracle Cloud Free Tier - 24GB RAM / 4 vCPU ARM instance, free forever

• SubflowAI - Substack Notes scheduling Chrome extension (built by Dheeraj)

• (The AI Maker) - Substack writer with in-depth Claude Code and ClawBot knowledge and content

• - Security-focused Substack writer to follow

• - Thanks for valuable advice when Dheeraj was near the finish line with SubflowAI

Don’t forget to follow & subscribe the host

The Sustainable Solopreneur

I help writers build one person business without burning out. Every week, get actionable tips on writing, AI systems, and building income around your life. No hustle-bro energy. 🌱

By Vinayak Ramesh

Get PluggedIn

Stop paying $200+ a month for tools Claude Code can replace this weekend.

See what's included →

Every month without the actual scripts is another $225 you didn't have to spend.

Get PluggedIn to go from knowing the Oracle + n8n migration is possible to having the scripts that make it done