Build Your Second Brain with an AI Co-pilot
In this entry
- A second brain does two jobs: logging and retrieving
- Why a folder of markdown files beats a PKM app
- How Claude Code and Claude Cowork work
- Setup
- 1. Create a local folder for your notes
- 2. Install Obsidian as your editor
- 3. Pick Cowork or Code
- 4. Add a starter CLAUDE.md
- 5. Set up a safety net
- Logging: collect knowledge with a few skills
- YAML metadata basics
- When metadata turns into a table
- Retrieving: ask Claude, then search and browse
- 1. Ask Claude
- 2. Search
- 3. Links and graph view
- 4. Canvas
- 5. Databases (Obsidian Bases)
- A worked example
- What you still do yourself
- Limits and privacy
- A first-week routine
- What's next
- Related
This is a guide to building your personal knowledge management (PKM) system with Claude Cowork or Claude Code. It's written for people who take notes but don't code: you're fine with Notion or Google Docs, and a terminal makes you nervous.
You want an AI that works on your notes, not just a chat window. This post shows how.
A second brain does two jobs: logging and retrieving
- Logging: get things in fast. Notes, ideas, links, meetings, journals. The AI lowers the cost of capture. You dump raw input, and it cleans up, tags, and files it.
- Retrieving: get things back out when you need them. Ask questions, find related notes, summarize a topic, draft from what you already wrote.
Most PKM advice spends too much time on the system in between: folders, tags, templates. With a co-pilot, you can keep that system simple and focus on these 2 jobs.
The AI doesn't replace your own thinking, though. Organizing and reviewing stay with you (more on that below).
Why a folder of markdown files beats a PKM app
There are two ways to keep your notes:
| PKM app (Notion, Evernote, etc.) | Markdown files (Obsidian, any editor) | |
|---|---|---|
| Where notes live | Inside the app's own database, usually in the cloud | Plain text files in a folder you own |
| Who can open them | Only that app fully understands the format | Any editor: Obsidian, VS Code, a plain text editor |
| Where the AI comes from | The app ships its own AI features | Any AI that can read a folder can work on your notes |
| Leaving | You export, and hope nothing breaks | Nothing to export. It's already yours |
| Getting started | Easier and more polished | Needs a little setup |
The trade-off is real. Apps are easier to start with. Files need some setup, but you get ownership and flexibility. Once AI can read and edit a folder for you, the setup cost drops and the flexibility matters a lot more.
How Claude Code and Claude Cowork work
Everything happens on your computer, around one folder of plain files. You read and write notes in Obsidian. Claude works on the same files, either from a terminal (Claude Code) or from the Claude desktop app (Cowork).
Hover or tap a box to see what it does.
- What it can see: the files in your folder.
- What it can do: read, create, and edit files, and run tasks for you.
- How it knows your rules: a project instructions file,
CLAUDE.md. - Skills: reusable recipes you teach it once (see Logging below).
- You stay in control: it asks before it acts, and you keep a safety net to undo changes (see Setup).
Setup
1. Create a local folder for your notes
2. Install Obsidian as your editor
3. Pick Cowork or Code
Both work on the same folder, with the same notes. The difference is mainly how you talk to Claude.
| Claude Cowork | Claude Code | |
|---|---|---|
| Where you use it | The Claude desktop app | A terminal (also in the desktop app and code editors) |
| How it feels | Chatting with an assistant that can open your folder | Typing commands in a text window |
| Best for | Everyday logging and retrieving | Building your own skills and automations |
| Learning curve | Low. Point it at a folder and start asking | Higher. You need to be OK with a terminal |
| Plan needed | Paid Claude plan | Paid Claude plan |
My advice: start with Cowork. Move to Code once you want more control, for example writing your own skills or running tasks on a schedule.
4. Add a starter CLAUDE.md
CLAUDE.md is a plain note at the top of your folder. Claude reads it at the start of every session, so it's where you write your house rules once instead of repeating them in every chat.
Start small. Here's one you can copy and adjust:
# My second brain
This folder is my personal notes, opened in Obsidian.
## No folders
All notes live in one folder. Don't create subfolders.
Metadata, not folders, says what a note is.
## Metadata
Every note starts with YAML metadata. Always set `type`:
- journal: one note per day, named YYYY-MM-DD.md
- literature: notes on books, articles, videos
- evergreen: my own ideas, in my words
- person, company: people and organizations I keep coming back to
- moc: index pages that link notes on one topic
Reuse existing field names (date, status, source, author, tags).
Don't invent a new field without asking me.
## Rules
- Write in plain English, short sentences.
- Link related notes with [[wikilinks]]. Only link to notes that exist.
- Never delete a note. Set `status: archived` instead.
- Ask me before editing more than 5 notes at once.
Notice there are no folders. Every note sits in one folder, and the type field does the sorting. A note never has to live in one place: a book note can also be part of a project, and you just add a tag or a link instead of deciding where to file it.
Treat it as a living document. When Claude does something you don't like, add a rule. When a rule stops mattering, remove it.
5. Set up a safety net
Logging: collect knowledge with a few skills
- Skill 1: download an article or YouTube video into a note. Creates a literature note: what you took from a source.
- Skill 2: process a book, long document, or research paper into raw sources, a summary of the key concepts, and how those concepts link together. Creates literature notes, evergreen notes (your own lasting ideas, in your words), and a MOC (Map of Content: an index page that links related notes on one topic).
- Skill 3: journal logging. Creates a journal note: your daily log.
YAML metadata basics
YAML metadata (Obsidian calls it "properties") is a small block of labeled facts at the top of a note, between two --- lines. The note body is for you to read. The metadata is for sorting and filtering: what kind of note it is, its status, its date, and what it links to.
You rarely write it by hand. Your skills tell Claude to fill it in when it creates a note.
A literature note for a book:
---
type: literature
source: book
title: "Deep Work"
author: "[[Cal Newport]]"
status: done
rating: 5
finished: 2026-10-12
tags: [focus, productivity]
---
A journal note:
---
type: journal
date: 2026-11-22
mood: good
people: ["[[Sarah]]", "[[Tom]]"]
---
An entity note for a person:
---
type: person
company: "[[Holistics]]"
role: Designer
last_contact: 2026-11-15
---
A few rules of thumb:
- Always set
type. It's the one field every note needs. It replaces folders, and most filters start with it. - Use the same field names everywhere.
statusin one note andstatein another breaks your filters. Put your field list inCLAUDE.mdso Claude stays consistent. - Put links in quotes.
"[[Cal Newport]]"is a link to Cal's entity note, so it shows up in backlinks and the graph.
When metadata turns into a table
Once your notes share the same fields, Obsidian can turn them into a table, like a spreadsheet built from your notes. Here are all notes with type: literature and source: book:
| Note | Author | Status | Rating | Finished |
|---|---|---|---|---|
| Deep Work | Cal Newport | done | 5 | 2026-10-12 |
| Building a Second Brain | Tiago Forte | done | 4 | 2026-09-03 |
| Thinking, Fast and Slow | Daniel Kahneman | reading | ||
| The Mom Test | Rob Fitzpatrick | to-read |
You never type this table. It's built from the metadata, so it updates whenever Claude adds a new book note. Change the filter, and the same notes answer different questions: "books I'm reading now", "people I haven't talked to in a month", "journal entries where mood was low".
How to build these tables is covered under Databases in Retrieving, below.
More on how I use each note type in my PKM best practices.
Retrieving: ask Claude, then search and browse
1. Ask Claude
This is the biggest change from a normal notes app. Search needs the right keyword, but with Claude you can just ask a question in your own words, and it reads your notes to answer.
Some prompts to try:
- Find: "What have I written about pricing?" or "Which books did I read on focus, and what did I take from each?"
- Summarize: "Summarize my journal for October. What kept coming up?"
- Connect: "How do my notes on Deep Work relate to my notes on Building a Second Brain? Where do they disagree?"
- Remember people: "What do I know about Sarah? When did we last talk, and what did I promise to follow up on?"
- Draft: "Draft an outline for a blog post on note-taking, using only my evergreen notes."
- Spot gaps: "Which topics do I have lots of literature notes on but no evergreen note of my own?"
One habit makes the answers much more useful: ask Claude to cite its sources. Add "link the notes you used" to the end of a prompt, and every claim comes with a [[wikilink]] you can click to check. If an answer has no source, treat it as Claude's guess, not your notes.
2. Search
3. Links and graph view
4. Canvas
5. Databases (Obsidian Bases)
Bases is a core Obsidian feature that turns metadata into tables (see the example above). You can build one by clicking through filters in Obsidian, or ask Claude: "Create a base of all my book notes, with author, status, rating, and finished date."
Claude saves it as a .base file, which looks roughly like this:
filters:
and:
- type == "literature"
- source == "book"
views:
- type: table
name: Books
order:
- file.name
- author
- status
- rating
- finished
A worked example
What you still do yourself
The AI is a co-pilot, not the pilot. It can file and link your notes, but it doesn't know what matters to you.
Reviewing is where the thinking happens. You notice patterns, drop what's stale, and decide what to act on. Skip it, and your vault becomes a pile the AI can search but you no longer understand.
My split:
- The AI proposes: tags, links, summaries, merges.
- I decide.
- I keep a weekly review habit, even a short one.
Limits and privacy
- Privacy: what gets sent to Anthropic when Claude reads your notes (e.g. your journal).
- Failure modes: over-tagging, links Claude invents to notes that don't exist, large edits you didn't review.
A first-week routine
What's next
This is the first post of a series. A short follow-up will cover Git for non-tech people.