A practical setup for AI-native operators. Stop re-explaining your company to every new chat. Start delegating without losing your voice.
Every conversation begins with Claude reading all four memory types. The system gets sharper every time you save a correction, confirm a judgment call, or update what is happening now.
Every time you open a new chat with Claude or ChatGPT, you start from zero. "Let me give you some context first." You explain who you are, what your company does, who is on your team, what tone you write in, what you are working on this week, what you are stuck on.
Then you ask the actual question.
This is the founder's hidden tax. It does not show up in your calendar. It shows up as the slow grinding feeling that AI is supposed to be making you faster, but somehow you are still the bottleneck. You are. Because all the context lives inside your head.
A Second Brain moves that context out of your head and into a place Claude can actually read it. Every chat starts pre-loaded with your company, your team, your tone, your principles, and the work in front of you right now.
Before: I open a chat. I explain. I ask. I correct the answer five times until it sounds like me.
After: I open a chat. I ask. The answer comes back sounding like me, with all my context already loaded.
This is not a productivity hack. It is a structural change in how you operate as a founder. Once it is in place, going back feels like writing software without version control.
The Second Brain is not a product you buy. It is a setup you build out of two pieces of software, neither of which is novel on its own. The leverage is in how they connect.
That is the whole stack. No new SaaS subscriptions. No API keys to manage. Your knowledge sits as plain text files on your laptop, and Claude reads them when you give it permission.
If you have a heavier setup (Claude Code, Cowork, a team subscription) it works the same way and gets more powerful. But this guide assumes you are starting from zero with just Claude Pro and Obsidian. That is enough.
Download and install Obsidian (it is free). Open it and create a new vault. Name it something you can live with for years, not something cute. Mine is just Second brain.
Inside the vault, you will create four folders and three files at the root. That is it for the structure. Resist the urge to add more on day one.
None of these files exist when you create the vault. You make them yourself. In Obsidian, right-click in the sidebar and pick "New folder" or "New note" to create each one. The folder tree above shows exactly what you should end up with.
This is loosely based on the PARA method (Projects, Areas, Resources, Archive) but stripped down to what an operator actually needs. Atlas holds the meta layer (the overview of overviews, the bird's-eye view of your whole operation), and crucially the file Today.md which is your live state. Projects holds anything that has a deadline. References is everything else that does not change much. Journal is your daily log.
Here is the rhythm that makes the whole system work.
Each day, you open a fresh Cowork chat with Claude (or a new conversation in whichever Claude product you use). You work normally. You think out loud. You ask questions, get answers, make decisions.
At the end of the day, you tell Claude what was worth keeping from the day. New information about a deal, a decision you made, a change in plans, anything you want tomorrow's chat to start knowing. Claude updates Atlas/Today.md for you.
Tomorrow morning, when you open a new chat, Claude reads Today.md first and the conversation starts already loaded with where you left off.
Morning: open new chat. Claude reads Today.md. You work.
Evening: you tell Claude what to save. Claude updates Today.md.
Repeat. The system gets sharper every day you run this loop.
At the root of your vault, create a file called CLAUDE.md. Think of it as the house rules for Claude when it works inside your vault. Folder conventions, tone preferences, what to read first, how to behave. Claude reads this file before doing anything else, so what you put here shapes every conversation that follows.
You can use the starter version below. Paste it into CLAUDE.md, then change "Write your name here" to your own name. Adjust the tone bullets to fit how you actually want to work.
CLAUDE.md / starter version# How to work in this vault
This is Write your name here's Second Brain. Content is in English
(or your language). Pure Markdown. No build steps. Wikilinks
([[Note Name]]) are used throughout.
## Conventions
- Always read Atlas/Today.md first before answering questions
- New project notes go in Projects/ AND get linked from Projects.md
- Every note starts with a # heading matching the filename
- Every project note has *Last updated: YYYY-MM-DD* near the top
- Cross-link liberally with [[wikilinks]]
- Save valuable conversation outputs as structured notes, not raw chat logs
## Tone
- Direct, honest, no corporate fluff
- Push back when you disagree
- Match my tempo - if I am moving fast, you move fast
- Warm but never sycophanticThat is the vault. Folders, files, conventions. The structure is done. From here on, you barely need to touch Markdown again. The next chapter shows how.
This is the move that separates a Second Brain from a fancy note app. The memory layer is a small collection of files that act as Claude's persistent memory across every conversation.
In the previous chapter you created four folders (Atlas, Projects, References, Journal). Now create one more, called memory/, at the root of your vault. This folder gets its own home because it works differently from the rest. The other folders hold your knowledge. The memory folder holds the rules Claude reads on every chat.
Before you set up anything, here is what lives in memory/. You do not need to write these files yourself, you give Claude a prompt and Claude builds them with you. But you need to know what the four types are so you understand what Claude is asking about.
Open Claude in your vault (Cowork mode, Claude Code, or just paste the vault path into a chat if you are on Claude Pro web). Then paste the prompt below.
Claude will interview you. Ten to fifteen minutes of conversation. By the end you have a memory folder full of files, a MEMORY.md index, and every future chat starts already knowing you. No Markdown writing required from your side.
After Claude is done, your folder structure should look like this:
Every memory file Claude creates will look something like this (so you know what to expect):
---
name: User role and background
description: My role, background, and how I want to be communicated with
type: user
---
I am the founder and CEO of [Company], building [Product] for
[market]. Background: [your two-sentence story].
Communication preferences:
- Direct and honest, push back when needed
- Plain language, no jargon without parentheses
- Founder mindset, not tech-only mindset
- Warm but never corporatememory/MEMORY.md is the index of all your memory files. One line per file. Claude reads it first on every conversation and uses it to know which deeper memory files to pull when relevant. Claude builds this for you as part of the interview prompt above. You do not need to maintain it by hand.
- [User role](user_role.md) - founder, not a developer
- [No em-dashes](feedback_no_em_dash.md) - never use em-dashes, AI-tell
- [Plain language](feedback_plain_language.md) - explain tech in parentheses
- [First customer goal](project_first_customer.md) - signed by end of June
- [Accounting is authoritative](reference_accounting.md) - live invoicing dataThe rule of thumb: keep the index under 200 lines. If it gets bigger, you are saving too much. Memory is for things that are surprising or non-obvious, not for things Claude could figure out by reading your projects. When in doubt, ask Claude "is this worth a memory file?" and let it judge.
Do not save sensitive personal data (addresses, ID numbers, passwords, health info) unless you absolutely have to. The vault is plain text on your machine. Anyone with access can read it.
How Claude reads the vault depends on which Claude product you use. There are three reasonable paths. Pick the one that fits where you already are.
This is the smoothest path if you are on macOS or Windows. The Claude desktop app has a feature called Cowork that gives Claude direct read and write access to a folder of your choosing. Point it at your vault folder. Claude will read your CLAUDE.md, your memory/MEMORY.md, your Atlas/Today.md, and any other file you reference automatically.
This is how we run this setup.
If you are comfortable in a terminal, Claude Code is a command-line tool that opens a Claude session inside any folder. cd into your vault, run claude, and you have a brain that knows your whole vault. Best for developers and operators who live in the terminal.
If you only have Claude Pro on the web, you can still make this work. At the start of any important chat, paste the contents of memory/MEMORY.md, Atlas/Today.md, and any project file you need. Less elegant, but it works.
If this is your path, build a simple "context starter" file you copy at the beginning of every session. It will not be as automatic as Path A or B, but it gives you 80% of the value.
The point is not the tool. The point is that Claude has a reliable way to read your context. Pick whichever path matches your current setup. Switch later if you outgrow it.
The final piece. Once Claude knows who you are and what you are working on, you can start delegating recurring jobs. An agent-file is a single Markdown file inside Projects/ that describes a job you want Claude to do on demand or on a schedule.
Like the memory layer, you do not need to write these by hand. You give Claude a prompt, Claude interviews you about the job, and Claude writes the file. From then on, when you want that job done, you reference the file by name.
Pick the most painful repeating job in your week. Sales reply monitoring. Weekly investor update. Cleaning up your inbox. Something you do over and over and would love to delegate. Then open Claude in your vault and paste this prompt.
Five to ten minutes per agent. Once it is written, you reference it by name whenever you want that job done. Claude reads the file and follows the rules.
Here is what a finished agent-file will look like after the interview. You can read it to understand the structure, or skip it - Claude handles the writing.
# Sales Outreach Agent
## Job
Monitor the Smartlead campaign and surface any reply that looks
qualified. Draft a response in my voice.
## When it runs
Daily at 11:00, on demand if I ask for "outreach status".
## Output
A short summary written to Atlas/Today.md under the heading
"Outreach status". Drafted responses go in Projects/Outreach-drafts/.
## Rules
- Never send anything automatically. Only draft.
- If a reply mentions pricing, escalate to me - do not respond.
- Match the tone of the original outreach: warm, founder, no SDR fluff.
- Flag anything that smells like a hot lead with [HOT] at the top.Start with one agent. Get it working for a week. Refine it. Only then build a second. People fail at this by trying to design ten agents on day one. After three months of one-at-a-time, you will have five to ten agent-files doing real work in your operation.
A Second Brain that does not learn is a wiki. The feedback loop is what makes it a brain.
Three habits do almost all the work.
The first time you correct Claude on something it could have known, ask Claude to save it as a feedback memory. Include the reason so Claude can judge edge cases later.
Wrong way: "Stop using em-dashes." Then forget. Hope it sticks.
Right way: "Save this as a feedback memory: never use em-dashes, it is an AI-tell. Apply everywhere, every chat." Claude writes the file. It sticks.
This is the move most people miss. When Claude makes a non-obvious judgment call and you confirm it ("yes, that was the right call"), tell Claude to save that too. Otherwise you only train it on mistakes and it becomes overly cautious over time.
At the end of each day, tell Claude what was worth saving. Decisions made, new information that came in, where you left off, where tomorrow needs to pick up. Claude updates Today.md. Tomorrow's chat opens with everything still loaded.
After 30 days of these three habits, you will notice the answers start to sound more like you than you expected. After 90 days, you will not be able to imagine working without it.
If you used the prompts in Chapters 4 and 6, Claude has already written files like these into your vault. The patterns below show you what the output looks like, so you know what you are getting. They also work as starter templates if you ever want to add a memory file by hand later.
---
name: Founder identity
description: Who I am, what I am building, how I want to be communicated with
type: user
---
I am the founder of [Company]. We build [one-line product description]
for [target market].
My background: [one or two sentences].
How I want to work:
- [Preference 1]
- [Preference 2]
- [Preference 3]---
name: Sales context
description: Current sales motion, ICP, deal stage, and what I am chasing
type: project
---
ICP: [Specific ICP description]
Primary channel: [How we reach them]
Stage: [Where most deals are right now]
Active deals: [List, or pointer to where they live]
Next milestone: [Goal + date]
Why this milestone matters: [The forcing function]---
name: Team operating principles
description: How my team works and what I want preserved in everything Claude produces
type: feedback
---
Team facts:
- We are [size] people
- [Key team member 1] handles [X]
- [Key team member 2] handles [Y]
When producing team-facing material (specs, kickoffs, comments):
- Language: [English/Swedish/etc]
- Tone: [Direct/warm/etc]
- Format: [Markdown/Slack-style/etc]
Reason: [Why this matters - past incident or strong preference]---
name: Voice and tone
description: How I write, what to avoid, what to lean into
type: feedback
---
My voice is:
- [Adjective 1]
- [Adjective 2]
- [Adjective 3]
Never use:
- Em-dashes (AI-tell)
- Corporate buzzwords (synergy, leverage, etc)
- Bullet points in conversational replies
Lean into:
- Concrete examples
- Short sentences when making a point
- Honest pushback when I am wrong---
name: External systems map
description: Where the real-time data lives outside the vault
type: reference
---
Source of truth for:
- Invoicing and accounting: [System name + URL]
- Customer data: [System name + URL]
- Sales pipeline: [System name + URL]
- Team comms: [System name + channel]
- Calendar: [System name]
How to apply: If I ask about live numbers (cash, MRR, pipeline),
point me to the right system. Do not estimate from the vault.MEMORY.md is in the path Claude can see. In Cowork or Claude Code, the path should be inside the folder you opened. If Claude still misses it, paste the contents of MEMORY.md at the start of your chat and say "use these as your active memory". Then add a feedback memory reminding Claude to check the index first.Two tools, four folders, four memory types, and three habits. Twenty dollars a month. Set it up in an afternoon. Run it for years.
If you build yours and want to compare notes, find me on LinkedIn. I read every DM.
Roger is built by Adthekey AB. We design operational systems for service businesses, engineered in Sweden.
Visit Adthekey →