Most of us save articles and never read them again. That is the exact problem an AI second brain solves.
I sat down with Wyndo for 17th episode of our One Shot Show session to dig into the one he built and calls his “LLM Wiki.” This AI second brain does more than store what you read and hear. It connects ideas across everything you have saved, automatically, using Claude Code.
By the end of this article you will know the exact three-layer structure, the core slash commands, and the guardrails that keep the system from turning into a noisy, expensive mess.
👋 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.
Table of Contents
What Is the Three-Layer Architecture Behind an AI Second Brain?
Obsidian, Notion, or Plain Markdown: Which Foundation Should You Use?
Why Your Notes App Is a Graveyard, Not a Brain
Every note-taking app has the same failure mode. You save something, tag it, maybe link it to one other note, and then it sits there forever.
As Wyndo put it during the session:
“On normal note apps, you file it, you tag it, you link it, you remember it exists, but it sits there with nothing that can make connections out of it.”
That line is the whole reason this topic matters right now. Andrej Karpathy’s post on building a second brain went viral because it named a problem everyone with a “read later” pile already feels.
Wyndo built his own version of that idea using Obsidian and Claude Code together, and what he showed me convinced me this is not a productivity toy. It’s a real system for turning passive reading into compounding knowledge.
A second brain built with Claude Code does more than file your notes away. It reads them, classifies them, and finds the connections between a podcast you saved last week and an essay you saved two years ago, without you lifting a finger.
What Is the Three-Layer Architecture Behind an AI Second Brain?
The system runs on three layers: Source, Generate, and Cross-link. Source is anything you feed in (articles, podcasts, tweets, book notes, or your own thoughts). Generate turns that raw input into a structured wiki page.
Cross-link is where Claude checks what already exists and connects the new page to it.
Wyndo laid this out clearly, and it’s the foundation everything else in this article builds on. Without this structure, you end up with a bigger pile of disconnected files, only now an AI wrote them for you instead of you copy-pasting.
The part that makes this work without burning through your Claude usage is the index. Instead of Claude reading every existing page in full to check for connections, it checks a lightweight log first. I put the risk of skipping this step bluntly:
“Otherwise, you will be just finishing your usage limits in a very good time.”
That’s not a throwaway line. If your second brain has 200 pages and Claude has to read all 200 in full every time you ingest one new article, you’re burning tokens on repetition instead of insight. The index is what makes this system scale past your first 20 notes.
What Commands Run the Whole System?
Wyndo’s second brain runs on five Claude Code slash commands, each with one job. These four do the daily heavy lifting:
Ingest URL - fetches an article, strips it to clean text, and generates a summary and key-points page
Process Inbox - handles fleeting thoughts you dropped in without formatting
Ingest Podcast - pulls in a podcast transcript and builds both an episode page and a guest entity page
Reflect - turns Claude into a Socratic dialogue partner that deepens a short personal aphorism into a fuller idea
Each command does one thing well instead of trying to be a universal “add to wiki” button. That’s a deliberate design choice. A single command that has to guess whether you’re feeding it a podcast, a tweet, or your own 3am thought ends up doing all three badly.
On the Reflect command specifically, Wyndo made a point I want to preserve exactly as he said it:
“AI doesn’t reflect your skill at prompting, it reflects your thinking. The quality comes from the refinement loop, not the model itself.”
That’s worth sitting with. The Reflect command isn’t magic. It’s a structured back-and-forth that forces you to clarify a half-formed idea, and the AI is only the mirror.
What Guardrails Stop an AI Wiki from Becoming a Noisy Mess?
Without limits, an AI that auto-links your notes will connect everything to everything, and the wiki becomes useless. Wyndo built four explicit guardrails to prevent that, and these are the most copyable part of the whole system.
Rule Breakdown for a Lean AI Wiki
Link cap: 5-8 links maximum per page, no exceptions
Person/podcast threshold: don’t create a dedicated page for a person or show until they’re referenced at least twice
Linking direction: forward-only, meaning Claude never goes back and retroactively edits old pages to add new links
Edit ceiling: each ingest can touch at most 2-5 existing pages
Here’s why each one matters. The link cap keeps any single page from becoming a hub that connects to unrelated ideas simply because a keyword matched. The twice-referenced rule stops you from getting a dedicated page for every random name mentioned in a podcast intro.
The forward-only rule keeps ingestion fast, since Claude isn’t re-scanning your entire archive every time. And the edit ceiling keeps token costs predictable, since you know the ceiling on how much any single command can touch.
If you’re building this yourself, don’t skip these rules to “see what happens.” Wyndo was direct about this: give the system a month or two to evolve rather than trying to perfect it on day one. The guardrails are what keep that evolution from turning into chaos.
Get PluggedIn
Want the shortcuts behind builds like this one?
Get PluggedIn unlocks the full GenAI Unplugged asset library: the command packs, templates, and Claude Code toolkits we build on the show and in every article.
If you’re wiring up a second brain, a triage system, or your own automation, it’s the fastest way to skip the trial and error and start from something that already works.
Obsidian, Notion, or Plain Markdown: Which Foundation Should You Use?
Obsidian stores every note as a native plain Markdown file, which is why Wyndo picked it over Notion. Notion structures content into blocks behind an API layer, and that extra layer between Claude and your real words hurts comprehension.
Markdown gives Claude direct, unfiltered access to what you wrote.
Wyndo also built an Obsidian CLI skill that Claude Code calls to generate graph views, databases, and canvases on top of the plain files. That’s the tooling-rich path: visual, plugin-driven, good if you like seeing your notes as a graph.
I run something different, and I want to be clear this isn’t the “better” answer, only a different tradeoff. My setup skips Obsidian and Notion entirely.
It’s plain markdown files pointed at a single directory, plus MCP (Model Context Protocol, a way to connect Claude to outside tools) connections to email, Substack, and YouTube for triage.
“I don’t need the visualization of how my thoughts are linked. I just need to point it to that one directory and it is going to adapt everything.”
If you like seeing your knowledge graph, go the Obsidian route. If you’d rather have zero apps between you and your files, mine works fine too. Both run on the same Source, Generate, Cross-link logic underneath.
What Does the System Produce? A Live Example
Wyndo ingested a real episode live: the Modern Wisdom podcast with Chris Williamson interviewing Mark Manson. One command produced a full entity page: summary, key points, notes, and notable quotes, all templated and ready to cross-link.
He then ingested a second episode, this one with David Friedberg. The system automatically linked the new page back to the existing Modern Wisdom and Chris Williamson pages it had already built, without Wyndo telling it to.
The most striking moment for me was when Paul Graham’s essay “When to Do What You Love” got auto cross-linked to Wyndo’s own personal note titled “Hard Work Is Overrated.” Neither page mentioned the other explicitly. Claude found the conceptual overlap on its own.
That’s the real payoff of this whole system: connections you wouldn’t have made by hand because you’d forgotten the older note even existed.
Do You Need a Separate Second Brain for a Specific Topic?
No. This came up as an audience question: someone wanted to go deep on one investment advisor’s content specifically. Wyndo’s answer was simple: don’t build a whole new system, add a dedicated page for that person inside the second brain you already have.
For a narrower use case, like wanting Claude to answer in that specific person’s voice and principles, he pointed to NotebookLM with a custom system prompt as a lighter-weight alternative. You don’t need infrastructure for every niche interest. You need one system with enough structure to hold all of them.
How I Organize My Own Knowledge Vault
I also walked through my own structure, which runs on Tiago Forte’s PARA method (Projects, Areas, Resources, Archive). My Knowledge Vault breaks into Inbox, Literature, Permanent, Projects, Daily, Maps, Output, and Systems, sitting alongside a Core folder and an Archive.
Wyndo organizes his project management the same way, which says something. Two people built genuinely different tools (Obsidian-based versus plain markdown) and landed on the same organizational backbone underneath. The specific app matters less than having a consistent place for things to land.
One closing note that applies no matter which system you build: slash commands still work, but they’re becoming legacy.
The newer approach is Claude Code Skills, and if you’re building a command that should only run when you explicitly call it (not something Claude decides to invoke on its own), set model invocable to false when you create it.
Frequently Asked Questions
What is an LLM Wiki / AI second brain?
It’s a system where Claude Code automatically ingests notes, articles, podcasts, and personal thoughts, then finds correlations between them and generates cross-linked wiki pages. It replaces static, manually tagged files with an active, connected structure.
What are the core layers of an AI second brain system?
Three layers: Source (raw input like articles, podcasts, tweets, personal thoughts), Generate (templated wiki pages with summaries and key points), and Cross-link (connecting the new page to existing related pages via a lightweight index).
What slash commands run a second brain like this?
Wyndo runs five slash commands in total. The four core ones: Ingest URL for articles, Process Inbox for fleeting thoughts, Ingest Podcast for transcripts and guest pages, and Reflect for a Socratic dialogue that deepens a short personal idea.
How do you stop an AI-linked wiki from becoming a mess?
Cap links per page at 5-8, don’t create a dedicated page for a person or podcast until it’s been referenced twice, link forward only, and limit each ingest to editing 2-5 existing pages.
Do you need Obsidian to build an AI second brain with Claude Code?
No. A plain markdown-only setup pointed at a single directory, paired with MCP connections to email and other tools, works without any note-taking app at all.
Should I build a separate second brain for one specific person or topic?
No. Add a dedicated page inside the second brain you already have, or use NotebookLM with a custom system prompt if you want responses in that person’s specific voice.
Key Takeaways
Build the index before you scale content. A lightweight log of existing ideas is what keeps Claude from re-reading your whole archive (and burning your usage limits) on every new ingest.
Cap links per page at 5-8. This one rule prevents your wiki from turning into a tangle of low-value connections.
Wait for two references before creating a page. A name mentioned once in passing doesn’t deserve its own entity page.
Link forward only, never backfill. This keeps ingestion fast and keeps token costs predictable per command.
Pick your foundation based on whether you want visualization. Obsidian gives you a graph view and plugins. Plain markdown gives you zero intermediary layers.
One second brain, not one per topic. Add a page for a narrow interest instead of spinning up a whole new system.
Give it a month or two before judging it. The guardrails are what let the system mature instead of collapsing under its own noise.
Resources Mentioned
Obsidian: Note-taking app storing plain Markdown natively, with a CLI Claude Code can call for graph views and canvases
Claude Code: The automation engine running the ingest, inbox, podcast, and reflect commands for both systems discussed
Notion: Compared unfavorably here because its structured-block layer sits between Claude and your raw notes
NotebookLM: Suggested for simulating one specific person’s voice and principles via a custom system prompt
Andrej Karpathy’s second brain post: The original viral post that sparked this wave of AI second-brain builds
Modern Wisdom (podcast): Demoed live during the session, hosted by Chris Williamson, guest Mark Manson
David Friedberg: Podcast guest whose episode was ingested live, auto-linking to the existing Modern Wisdom pages
Paul Graham: His essay “When to Do What You Love” auto cross-linked to Wyndo’s personal note during the demo
Tiago Forte / PARA method: The organizational framework both of us use for our knowledge vaults
MCP (Model Context Protocol): What connects my own setup to email, Substack, and YouTube for triage
Your 15-Minute Challenge
Pick one article you saved this week and never read again. Open Claude Code, point it at a single folder, and ask it to summarize the article and list its key points into a new markdown file.
That’s the entire Generate layer in miniature. Success looks like one clean page with a summary and key points you can reference later.
Once that feels natural, add the cross-link step: point Claude at your growing folder and ask it to connect each new page to the two or three most related notes you already have. That’s how one page becomes a system.
