Wyndo asked me the question every solo creator gets: “How does one person ship this much?” This is the full written breakdown of what I showed on screen. The numbers, the honest parts, and what breaks at 2am.
As of May 2026, I run GenAI Unplugged and Discover With Dheeraj solo. Three articles a week. Two YouTube channels. An in-person Academy. Twelve-plus pieces of shipped software in the last few months. No team. No hand-written code in eight years.
When Wyndo asked how one person does all of this, my honest answer was: they don’t. Not manually. The stack does it.
But “I have a stack” isn’t useful to anyone.
What’s useful is the map: which layers exist, what each one does, where they talk to each other, and what breaks when you get it wrong.
By the end, you’ll know which layer to build first and which to leave for later.
👋 Julley, I'm Dheeraj, an AI systems builder.
I build production-grade AI systems at work by day and ship my own products by night, 9 and counting, including SubflowAI and the Content OS Agents Toolkit. This newsletter is the bridge between those two worlds. Every system, every build, documented step by step.
Join 1,800+ builders getting the exact AI setups, prompts, and workflows that actually work in your business.
The stack doesn’t make you faster. It makes the work possible at scale without a team. And the parts that break will teach you more than the parts that work.
What Are the 6 Layers of a Solo Creator’s AI Stack?
Six layers. One orchestrator in the middle. Everything else is a spoke.
Layer 1: The Brain: Claude Code + Codex
Layer 2: Research: Gemini → Tavily → Perplexity
Layer 3: Creatives & Video: GPT-4o Images, Kling, Veo, Descript, Remotion
Layer 4: Automation & Browser: n8n, Playwright, Chrome Extensions
Layer 5: Data: Notion, SQLite, Substack Intelligence
Layer 6: Intelligence: Skills, Sub-Agents, Manthan
When all six layers work together, the output is: 12+ products shipped, 3 articles a week, a YouTube channel, an Academy, and a sane Saturday.
Let’s walk each one.
What Tools Power the Brain Layer of My AI Stack?
Claude Code reads files, runs commands, calls APIs, and builds software. That difference matters.
The brain layer has a few pieces
CLAUDE.md is a system file. It gives Claude Code everything about my business: who my audience is, what my products are, what my voice sounds like, and what I never want to do. It’s business memory. When you count all the connected files, it’s 4,000 lines.
On top of that, I have 50+ custom slash commands. Each one is a markdown file that encodes a workflow. /draft-post runs my entire 9-stage content pipeline. /yt-research starts the YouTube research agent I built with Claude Code. /hero-image generates and places a Substack cover image for a queued article. One command. Six layers of context. One output.
Codex sits behind Claude Code as my backup brain. Same system files, same commands, different vendor. If Claude is rate-limited or down, I switch to Codex without losing anything. I’m never one outage away from a stopped business. That’s architecture, not paranoia. It’s how you build systems that hold up.
MCP servers extend the brain outward.
Notion MCP, Substack MCP, Tavily MCP, a custom archive MCP for my newsletter database. The agent can talk to my real tools, not only my files.
I also have 30+ specialist sub-agents, each with a single job. SEO researcher. Brand voice auditor. Fact-verifier. Content gap analyzer.
The single-responsibility principle applies to agents the same way it applies to functions. The moment an agent tries to do two jobs, it gets worse at both.
One question Wyndo asked on the show is worth repeating: “Claude Code vs ChatGPT, what’s the real difference?“
Two things.
First, context stays. My CLAUDE.md and project files give Claude Code lasting business knowledge. ChatGPT loses that context every conversation.
Second, deep tool access through MCP. It can query my Notion database, my SQLite archive, my YouTube Analytics API. The difference between asking for help and giving someone the keys.
Commercial break: Claude Code Builder cohort
The founding batch of my Claude Code cohort starts in July on Maven. Four weeks, six sessions. Learn to build the AI system with me.
Only 12 Seats. When they’re gone, the founding price ($797) closes and Cohort 2 opens at $1,597.
Use code GENAI20 for 20% off. Expires July 9. Check the Syllabus →
How Does the Research Layer Stay Free While Covering Everything?
Three tools. One layer. Free first, paid last.
Gemini (via CLI) handles quick lookups for free. I call it via the
gemini-searchskill when Claude’s built-in web search doesn’t do the job. Free Google Cloud credit. No billing.Tavily is the primary research tool. Search, crawl, map, deep extract, all in one MCP. 1,000 free credits a month. Most weeks I don’t hit the limit.
Perplexity is the backup. It only fires for complex research questions, or when Tavily is down or out of credits. It’s the only paid tier in this layer.
Why a cascade instead of paying for the best tier outright? Two reasons.
Cost: my content pipeline does dozens of searches per article. If every search hits the paid tier, that’s $20-30/month on research alone. The cascade keeps that cost near zero.
Resilience: I learned this the hard way. Firecrawl used to be in this slot. They rate-limited me last April and my whole research pipeline broke until I swapped to Tavily. Now I keep Perplexity behind Tavily. If one provider breaks, the pipeline doesn’t stop. Never one vendor away from a stopped business.
The pattern repeats across every layer. Redundancy is not a luxury. It’s the only way a solo operator avoids firefighting at 2am.
How Does the Creatives Layer Turn Visuals Into a Pipeline?
Visuals are pipelines, not folders. This is the layer that took me longest to wire correctly.
GPT-2 Images is my primary image model right now. Two reasons: character consistency and infographic accuracy. For work that needs a consistent character, GPT-2 Images holds the face and outfit better than what I’ve tested. For infographics, when I give it specific layout instructions, it follows them more precisely.
I use the Batch API for hero image generation. That cuts cost by 50%. The images generate overnight, land in the project folder with the right filename, and I never have to open a file picker.
Gemini Nano Banana 2 backs up image generation when GPT 2 Images credits run low.
Video Editing AI Tools
Kling handles animated covers. Substack thumbnails, article headers. When I need motion without full video generation, Kling is the faster and cheaper option.
Veo handles video clips that go inside longer videos. It costs more than Kling, so the split is on purpose. Kling for covers. Veo for in-video clips.
Descript is my video editor. It connects directly to Claude Code via an MCP server. Descript has an internal AI agent called Underlord. When I connect Descript’s MCP, Claude Code can send instructions to Underlord.
Instructions like: “clean up filler words, remove silences, fix audio levels.” I stay in my terminal. Descript does the edit but it is very frustrating and mostly not that good. I just use because it has script based video edit, AI features are bad for my workflow at least and it is costly too.
I’m building my own lightweight video editor using Claude Code to replace Descript for this workflow. But for now, Descript does the job.
Remotion is the interesting one. It renders animated overlays as clear video files. Title cards, text highlights, transitions. Claude Code writes the React component (a UI building block). Remotion renders it.
ffmpeg (a video processing tool) combines it with the source footage. The output is a finished video with automatically generated B-roll. A fully animated overlay that tracks to the narration. All of it runs inside my Claude Code terminal.
Two full weekends. That’s what building a solo AI stack without proven architecture costs you. Not in money, but in the kind of Saturday you don’t get back. The demo took one weekend. Fixing what the demo didn’t cover took another. Most people hit that second weekend unprepared.
If you’re building something real on top of AI tools, the patterns that prevent that second weekend exist. They’re what PluggedIn is built around.
How Does the Automation Layer Move Data Between Tools Without Manual Work?
n8n is the plumbing. Self-hosted on Oracle Cloud’s always-free tier. It costs zero per month. It moves data between layers.
Hero images get generated, n8n routes them to the right post folder. An article finishes the pipeline, n8n syncs the state to Notion. A live show ends, n8n processes the clips and sends them to LinkedIn.
I haven’t manually logged into LinkedIn in months. The clips you see from One Shot Show on my LinkedIn feed come from n8n. Substack gives me the clips. I run /process-clips. They upload to my Oracle server. n8n handles the rest.
The other point worth noting about n8n vs Claude Code
Claude Code manages the instance. I don’t SSH in to fix failed workflows. I don’t restart workers. When a workflow fails, the error goes into the run log. Claude Code reads that log, tries to fix the failed node, and re-runs it.
If the second attempt fails, it stops and flags it for me in the morning. The server manages itself.
For a full breakdown of how n8n and Claude Code compare across eight real categories, that article covers when each tool handles which job in the stack.
Playwright handles browser automation when no API exists. LinkedIn posting is one example. Some Substack features don’t have APIs yet, so Playwright fills the gap.
Chrome Extensions are where automation becomes products. SubflowAI is a Notes scheduling and deep analytics tool for Substack. It’s a Chrome extension. 80+ paying users. $69 lifetime.
Built entirely through Claude Code. No hand-written code. My LinkedIn engagement extension and Substack engagement manager follow the same pattern.
What Makes Data the Memory of the AI Stack?
Data is the memory of the system.
Notion is where the calendar lives. Every article in production, every YouTube video planned, every collab tracked. Notion is the structured source of truth. The Notion MCP means Claude Code can query and update the calendar directly without a browser.
SQLite is my local archive. A database with 100+ published articles, engagement data, pillar tags, and series metadata. Why SQLite instead of flat files? Fast text search.
I built an archive MCP server on top of this database. Claude Code can now query the archive: “what have I written about n8n in the last 6 months” returns results in milliseconds.
SQLite also handles subscriber sign-in. When someone buys access to a tool, I create a magic link tied to their Substack email. I check it against the subscribers table in SQLite and grant or deny access.
Zero calls to Substack’s API at runtime. The sign-in layer is a single database query. The PluggedIn MCP built on top of that subscriber database is how that authentication layer became a live product.
Substack Intelligence via Substack MCP Server
Substack Intelligence is a custom sync layer. It pulls all my subscriber and post data into SQLite. Subscriber groups, engagement rates, inactive users, revenue breakdowns, all searchable through MCP.
I run the sync manually in the mornings. The intelligence layer reads SQLite only. No API calls during analysis. How the Substack analytics MCP server came together is the full origin story of that build.
Notion vs. Obsidian
The Notion vs. Obsidian debate came up on the show. Wyndo is fully on Obsidian. I use both, but rely on Notion for anything structured or shared. Obsidian is better for personal knowledge bases (PKM, personal knowledge management). Notion is better when the data has tables and relationships, and needs to live somewhere your collaborators can reach it.
I will say I love notion is
Structured Database, I love tables still in this world of JSON,
Mobile App Access which I can use to approve, reject, review lot of content sitting in my son’s class or in boring parties.
I can publish Notion Pages publically for my readers/viewers who want to read the guides, access the files. Those are free website hosted pages literally without any paid subscription or maintenance. You can even publish all guides, lead magnets through it.
If you want to collaborate with others or Team members, then Notion has the edge.
It used to take lot of tokens, that I solved using the API directly. However, literally today/yesterday they have launched the Notion CLI to directly integrate itself into Claude Code to make it more efficient.
All Free tier of Notion.
Obsidian being local and have md files have its own advantage if the workflow works for you that ways. It is much lighter.
How Does the Intelligence Layer Make the Stack Smarter Over Time?
Claude Skills are reusable capability packs. Each skill is a folder with a SKILL.md file plus supporting assets. The hero image skill knows my brand visual system. The LinkedIn carousel skill knows the template dimensions.
When Claude Code loads a skill, it gets all of that context in one shot. How Claude skills work as reusable, portable capability packs goes deep on building your own.
I have around 90 skills at this point. Most took an afternoon to build.
Sub-agents are specialists. Each one has a single job: input goes in, output comes out. Three examples:
Fact-verifier: checks claims against live documentation
Brand voice auditor: catches passive voice and banned phrases
SEO researcher: pulls real SERP data
They don’t work with each other in real time. They hand off files. Bard writes avatar-script.json. Voicesmith reads avatar-script.json. No shared state. Clean handoffs.
When does multi-agent stop helping and start adding orchestration tax? The line is: when two agents need to work together live, you’ve crossed it. The fix is always to make the handoff a file. If you can replace a multi-agent system with a chain of single-purpose tools, you should.
Manthan is my video production system. Eleven specialist agents across three tiers. Tier 0 is a one-time channel setup agent. Tier 1 is the per-episode production chain and Tier 2 is Strategist.
It’s a daily analytics agent that reads YouTube data, detects what’s working, and writes winning patterns back into the channel config. The next episode picks up those improvements automatically. The system gets better at making videos without me changing anything.
What Actually Breaks in a Solo AI Stack and How Do You Recover?
No AI stack tour is complete without the honest section.
Debugging is painful.
When n8n fails without a clear error, you’re clicking through 15 nodes to find where the data went wrong. No-code doesn’t mean no-thinking. Sometimes it means more thinking, because the tools hide what’s happening underneath.
You don’t own the ground.
I built SubflowAI on top of Substack’s DOM structure (the underlying HTML layout of their web app). Substack updates their UI and my Chrome extension needs an emergency patch. Firecrawl rate-limited me and my research pipeline broke.
Kling’s pricing changed. Veo had a model update that changed output quality. You’re building on rented land. When the landlord remodels, you move furniture at 2am.
The 80/20 trap is real.
AI gets you 80% of the way in 20% of the time. The demo works. It looks incredible. And that’s what gets people into trouble.
The last 20% is where 80% of the real work lives. Security. Performance. Error handling. Edge cases. “What if the input is blank.” “What if two users submit at the same time.”
I shipped SubflowAI happy with it. I tested everything. Everything worked. Within days of going public, someone ran a denial-of-service attack and the whole system crashed.
I had to build a rate-limiting layer. I had to harden the auth. That took another full weekend. The weekend to build the app, and a weekend to make it not fall over.
That’s the journey. Demo to Dependable is a real series for a reason.
What I tell every cohort: start small. One CLAUDE.md. One slash command. One weekend.
The map isn’t the moat. Anyone can copy the diagram. What you can’t copy is the year of swapping tools when one broke. The hundred failed experiments that taught me which layers need to talk to each other. The architecture is free. The judgment is earned.
Key Takeaways
A CLAUDE.md file is the foundation of the entire stack. Every agent, skill, and slash command reads from it. Build this before anything else.
Three-tool research cascades (free → credits → paid) cut research costs to near zero. They also keep the pipeline running if any one vendor goes down.
The 80/20 rule works against you in production. The last 20%: security, rate-limiting, error handling. It takes 80% of the real time. Plan for it.
Single-responsibility agents beat generalist ones. The moment an agent tries to do two jobs, it gets worse at both.
File-based handoffs beat live agent coordination. If two agents need to talk in real time, you’ve added orchestration tax. Replace it with a handoff file.
Redundancy is not a luxury for solo operators. One vendor away from a stopped business is one outage away from a bad night.
Start with one command and one output. The rest follows naturally.
What’s the One Part of This Stack You Can Copy in One Weekend?
If you’re not a developer and you want one piece of this stack that works right now: write a CLAUDE.md file for your work.
Open a text file. Write down:
Who your audience is
What you’re building right now
What your brand voice sounds like
Three phrases you never want to use
Your three main topics
Save it as CLAUDE.md in the folder where you work. Start every Claude Code session in that folder.
Everything else in this stack builds on top of that file. The slash commands reference it. The sub-agents read it. The skills get shaped by it. If you skip this step and build everything else, you’ll come back and realize the whole thing needed this at the bottom.
Ten minutes. Five fields. That’s your start. Add more as you learn.
Frequently Asked Questions
Do I need to know how to code to use Claude Code?
No. I haven’t written hand-coded anything in eight years. Claude Code reads your files, runs terminal commands, and calls APIs on its own. What you need is the ability to describe what you want clearly. A CLAUDE.md that explains your business context is more valuable than any programming background.
How much does a stack like this cost per month?
Less than most people assume. n8n self-hosted on Oracle Cloud free tier: $0. Tavily research: $0 within 1,000 free credits (enough for most weeks). Claude Code: Max plan subscription.
GPT-2 Images Batch API: ~50% less than standard with overnight generation. The only consistently paid tier in my research layer is Perplexity, and it rarely fires. Most months the total additional API spend is under $50.
What is CLAUDE.md and how do I write one?
CLAUDE.md is a plain text file that gives Claude Code lasting business context. It sits in your working folder and gets read at the start of every session. Write your audience, your products, your brand voice, three phrases you never use, and your current projects.
That’s enough to start. Add more fields as you learn what Claude Code needs to know about your work.
What breaks first when you build a solo AI workflow?
Vendor dependency. I’ve had Firecrawl rate-limit me and take down my entire research pipeline. I’ve had Substack update their DOM structure and break SubflowAI overnight.
Every tool you build on is rented land. The fix is a fallback for every critical layer: two research tools, two AI providers, self-hosted infrastructure for anything you can’t afford to lose.
How is Claude Code different from ChatGPT for solo creators?
Two things. First, context stays. CLAUDE.md gives Claude Code lasting business knowledge. ChatGPT loses context every session. Second, deep tool access through MCP.
Claude Code can query your Notion database, your SQLite archive, your YouTube Analytics API directly. The difference between asking for help and giving someone the keys to your actual systems.
Related reading: Chain-of-thought and reasoning prompts, explained
