Section 1
The Moat Nobody Is Building
Every business on planet earth has data. Call transcripts. Emails. Slack messages. Meeting notes. SOPs. Client briefs. Old proposals. Years of decisions buried in inboxes.
Almost none of them are using it.
Andrej Karpathy posted the concept of a personal LLM wiki and 41,000 people bookmarked it. Maybe 100 of them will build it. By the end of 2027, every serious company will have one. The ones that move now own the compounding advantage. The ones that wait pay someone to build it for them.
That someone could be you. More on that in Section 6.
This guide gives you both paths to build a real one (the local one and the cloud one) plus the prompts that make either one actually useful.
Section 2
The Mental Model
Every second brain that actually works has the same shape. Three layers, three operations, one division of labor.
Three layers
Raw sources (read-only)
Everything you dump in. Meeting notes, clipped articles, screenshots, transcripts, briefs. The AI never edits this. You curate it.
The wiki (AI-owned)
Organized articles the agent builds from the raw layer. Linked, summarized, deduped, tagged. You never edit this by hand.
Schema and tags
The internal structure the agent maintains so retrieval is low-friction. You never see this layer. It just makes queries fast and accurate.
Three operations
Ingest. You hand it a file, link, or pasted text. It summarizes, files it as a new source, and updates the wiki pages it touches.
Query. You ask a question. It searches the wiki, synthesizes an answer, cites the sources.
Lint. Once a month you tell it to audit the wiki. It flags contradictions, sources older than 90 days, oversized files, and dead-end articles.
The division of labor
You curate the sources. The agent organizes, summarizes, retrieves. That is the deal. If you try to organize the wiki yourself, you have rebuilt Notion. If the agent tries to curate, it pulls in noise. Stay on your side of the line.
Section 3
Pick Your Path
Two real options. Both work. Pick based on how you actually want to use it.
Path A
Obsidian + Claude Code
- Cost: $20/mo (Claude Pro) or API metered
- Setup: 30 minutes
- Lives on: Your laptop
- You can: Read the files yourself, version it with git, edit anything in plain markdown
- Best for: People who write a lot, want full control, and like seeing the files
Path B
Hermes Agent + LLM Wiki
- Cost: ~$9/mo VPS + model tokens
- Setup: 10 minutes, one-click deploy
- Lives on: A cheap cloud server
- You can: Chat with it on Telegram from anywhere, get scheduled cron-job digests, run it 24/7
- Best for: People who want to query it from their phone and never look at files
Still not sure? Default to Path B. It is faster to ship, cheaper long-term, and you can always layer Obsidian on top later by syncing the files down to your machine.
Section 4
Path A: Obsidian + Claude Code (30 Min)
Two ways to do Path A. Both ship a working second brain.
Option 1. If you already have an Obsidian vault, the cleanest setup is the one-prompt scaffold I cover in the 30-minute Karpathy-method guide. Three folders, a CLAUDE.md template, and one prompt that builds the wiki from your raw dump.
Option 2. If you want the minimal version, do this:
Make the folder
Anywhere on your computer, create second-brain/. Inside it: raw/, wiki/, outputs/, plus a CLAUDE.md file.
Fill in CLAUDE.md
Paste the template below. Replace the bracketed parts with your real specifics. This is the file that makes a generic AI into your AI.
Dump 30 things into raw/
Meeting notes, client briefs, screenshots, articles, transcripts. Do not rename anything. Do not organize. Set a 15-minute timer and drop files.
Open Claude Code in that folder
Run claude in the terminal from inside second-brain/. Tell it: "Read CLAUDE.md, read every file in raw/, then build the wiki/ folder based on the rules." Walk away.
Open the folder in Obsidian
Point Obsidian at second-brain/ as a vault. Now you can read everything Claude built, follow links between wiki pages, and edit by hand if you want.
The CLAUDE.md template
Copy into CLAUDE.md
You are my AI thinking partner. This folder is your long-term memory. ## Context about me Name: [YOUR NAME] What I do: [ONE SENTENCE, e.g., "I run a 6-person consulting firm helping SMBs deploy AI"] Who I serve: [YOUR ICP] What I am optimizing for: [REVENUE / CLARITY / TIME, pick one] ## How this folder works - raw/ holds everything I dump. Meeting notes, article clips, screenshots, briefs, transcripts. Never organized by hand. - wiki/ is your territory. Read raw/, identify topics, build organized articles, link them together, maintain INDEX.md. - outputs/ is every answer you generate. Briefings, comparisons, decisions. These feed future answers. ## Your rules 1. Read CLAUDE.md, then wiki/INDEX.md, before answering any query. 2. When new files land in raw/, summarize, tag, and update the matching wiki article. Create new articles when needed. 3. Cite sources by raw/ filename when answering, so I can always trace the claim. 4. Once a month I will say "audit the brain". Flag contradictions, sources older than 90 days, oversized files, and orphan articles. 5. Never edit raw/. It is read-only. When I open a session, greet me and tell me what is new since the last time we talked.
Section 5
Path B: Hermes Agent + LLM Wiki (10 Min)
Hermes ships with a built-in LLM Wiki skill. That is the entire second brain function, already installed. Your job is to deploy the agent, connect it to Telegram, and feed it the first source.
Get a VPS
A cheap virtual private server is cheaper than a Mac mini, always on, and your data stays private. Any provider with a one-click Hermes template works. The KVM2 tier gives you headroom for multiple agents and runs around $9 a month on a long plan.
One-click deploy Hermes
In the VPS dashboard, find the Hermes Agent app template. Set an admin username and a password you will not forget. Click deploy. Wait two minutes for the green check. No terminal needed.
Connect a model provider
Sign into Hermes, pick quick setup, choose your provider. OpenAI Codex is the simplest if you already have a ChatGPT plan. OpenRouter is the cheapest if you want token transparency and access to free models like NVIDIA's NemoTron. Pick a default model. GPT-5.4 is a solid balance of cost and quality.
Wire it to Telegram
In Telegram, search @BotFather. Type /newbot. Name it whatever you want, get the bot token, paste it back into the Hermes setup. Open a DM with your new bot and run /set_home to lock that thread as your default channel.
Install the MarkDownload Chrome extension
This is the cheat code. Any article, tweet, or webpage becomes clean markdown with one click. Drag-and-drop the file into your Telegram chat with the bot and it becomes a new source.
Seed the wiki
Paste the prompt below into your Telegram chat with Hermes. Fill in the bracketed bits. Hermes will confirm the wiki is seeded and ask for your first source. Send it a MarkDownload-clipped article. Query it. You are live.
The Hermes seed prompt
Paste into your Telegram chat with Hermes
You are setting up a second brain for me from scratch using your built-in LLM Wiki skill. Domain: [WHAT THIS BRAIN IS FOR. e.g., "running my marketing agency" or "everything I'm learning about AI"] Owner: [YOUR NAME] Default storage path: use your default Operations you will run on this wiki: 1. Ingest. When I send you a file, URL, or pasted text, summarize it, file it as a new source, and update related wiki pages. 2. Query. When I ask a question, search the wiki first, then synthesize an answer that cites the source pages. 3. Lint. Once a month I will say "lint the wiki" and you will flag contradictions, stale sources older than 90 days, sources longer than 200 lines that should be split, and orphaned pages with no inbound links. Confirm the wiki is seeded and ready. Then ask me what I want to feed it first.
Your knowledge base will be the dumbest it ever is on day one. Keep feeding it. By day 90 it is a compounding asset you can query like a search engine that actually knows your business.
Section 6
Sell This As A Service
This is the part of the guide most people skip. Don't.
Every business owner reading this knows they should build a second brain. Most of them will not do it. They will not because they do not know what a VPS is, they do not have time, and they do not trust themselves to set up an AI agent without breaking something.
Charge them to do it. This is the AI equivalent of picks and shovels during the gold rush. The agents are getting cheaper and smarter every month. The data those agents need is sitting in inboxes, drives, and meeting notes that nobody has organized. Someone has to do the work.
The wedge offer
Second Brain Setup, $1,500 to $5,000 one-time
You build their second brain in a week. Deploy the agent, audit their data sources, seed the wiki, train their team to feed it. The high end is for businesses with messy data and 5+ employees who need access.
Curation Retainer, $500 to $2,000/mo
You run the monthly lint. You add new sources. You build new queries when their priorities shift. You sit in their Slack and update the brain when decisions happen. This is the recurring revenue play.
The 4-step delivery
Day 1, source audit. 60-minute call. Map every place their data lives: Gmail, Drive, Slack, Notion, CRM, recorded calls. Decide what gets ingested first.
Day 2 to 3, deploy and seed. Stand up Hermes on their VPS. Wire it to their team's Slack or Telegram. Ingest the first 50 sources. Build the schema around their business.
Day 4 to 5, train the team. Two 30-minute sessions. How to add a source. How to query. The five meta-prompts (Section 7). Send the team home with a one-page cheat sheet.
Day 30 / 60 / 90 reviews. Monthly check-ins. Run the lint, show them what the brain now knows that it did not a month ago, add new sources. This is where the retainer earns its keep.
Who to sell this to first
Small businesses between $500K and $5M ARR. They have enough data to matter, enough revenue to pay, and no internal team to build it themselves. Founders who are doing the work of three people. Anyone who has ever said "I need a chief of staff but I cannot afford one."
You are not selling AI. You are selling the moat their competitors are not building.
Section 7
The Meta-Prompts
A second brain you never query is just a folder. These are the five questions that turn it into a thinking partner. Ask them weekly. Ship the answers as outputs.
The procrastination flush
"Based on everything in this brain, what am I procrastinating? Be specific. Name the project, the next step I have been avoiding, and why I am probably avoiding it."
The compounding-task spotter
"Look at the last 30 days of my notes. What is one task I am doing manually every week that should be automated as a cron job or skill? Propose the automation and the prompt that would build it."
The decision support
"I need to make a call on [X]. Pull every relevant source, summarize what I have already decided on similar questions, and recommend an answer with the trade-offs spelled out."
The blind spot
"Read the wiki and tell me one thing I should know but probably do not. Either a contradiction between two of my sources, or a topic everyone in my industry is talking about that I have nothing on."
The tool builder
"What is one tool you could build me tonight that would make my work tomorrow noticeably easier? Build it."
Section 8
The Full Stack
Everything either path uses, in one place.
Obsidian
Local vault, free, you can read the files
Free
Claude Code
The agent that reads, writes, and queries the vault
$20/mo Pro or pay-as-you-go API
Hermes Agent
Built-in LLM Wiki skill, Telegram chat, always-on
~$9/mo VPS
MarkDownload (Chrome)
Clip any webpage to clean markdown in one click
Free
Telegram
Talk to your brain from your phone
Free
Section 9
What This Looks Like At Day 90
Day 1 it answers one question off one source. Useful, not magic.
Day 30 it has 100 sources, a schema that fits your business, and it can pull the receipts on any decision you have made in the last month.
Day 90 it knows your clients better than your CRM. It catches contradictions in your strategy before you ship them. It writes briefings for prospect calls in the time it takes you to make coffee. It tells you what you are procrastinating before you ask.
This is the compounding asset everyone is going to have by 2027. The window to build it before it is table stakes closes fast.
Start today. Pick a path. Drop the first source in.