Obsidian + Claude Code: Build Your AI Second Brain, Nishant Gupta

Open a new Claude chat and it knows nothing about yesterday. Not your project, not the decision you made last week, not the three times you already explained your own preferences. Every session starts at zero.

Obsidian, on its own, has the opposite problem. It is a folder of plain Markdown files sitting on your disk, doing exactly nothing until you open it and type. Beautiful, private, permanent. And completely passive.

Put them together and each one fixes what the other is missing. Claude gets a memory that survives between sessions. Obsidian gets a collaborator that reads, writes, and links on its own. This is that setup, step by step.

A chat window remembers nothing. A vault remembers everything.

01. Why Claude keeps forgetting you

Context windows are not memory

A context window is not memory. It is a workspace, and it clears the moment the session ends. Ask Claude the same question tomorrow and you are re-explaining your goals, your project, your preferences, from nothing, again.

That is fine for a one-off task. It falls apart for anything that compounds: a book you are writing, a business you are running, a body of research you are building. Progress needs a place to live between sessions, somewhere Claude can read on the way in and write on the way out.

That place does not need to be a database or a paid memory feature. It can be a folder of text files you already own.

02. Obsidian is just markdown

No lock-in, that is the whole feature

Obsidian is a local note-taking app built on one decision that matters more than any feature: every note is a plain .md file on your own disk. No proprietary database, no server round-trip, nothing that breaks if the company disappears. The core app is free. Sync and Publish are paid add-ons if you want them, the vault itself never requires them.

Because notes are just text, any tool that can read and write files can work with your vault, including Claude Code running from a terminal. You do not need Obsidian's own AI features or a plugin marketplace account. You need a folder, and something that can read it.

Same file, two readers. Neither one owns the format.

03. The five-folder vault

Structure is what makes it readable by a model

Open Obsidian, create a new vault, pick a folder. Anywhere is fine, ~/Obsidian/second-brain works. The structure matters more than the location, because a flat folder of five hundred notes is unreadable to you and to Claude both.

00 Inbox. Everything lands here first: a link, a voice memo transcript, a half-formed thought. Unsorted by design.

01 Daily. One note per day. The braindump, the log, the running record of what happened.

02 Projects. Active work only. A note per project, linked out to whatever it touches.

03 Knowledge. Evergreen notes. Things you have actually settled on, not raw captures.

04 Archives. Closed projects, dead ends, anything finished. Out of the way, never deleted.

00 Inbox

raw captures

01 Daily

daily notes

02 Projects

active work

03 Knowledge

evergreen notes

04 Archives

closed out

CLAUDE.md

root of the vault

04. Wiring Claude to the vault

Three ways in, pick one

Terminal, the simplest. Open a terminal, cd into your vault folder, run claude. Claude Code now reads and writes that folder directly. No config file, no extension, nothing to keep updated.

Claude Desktop, the GUI way. Click the plus icon in the message box, open Connectors, Manage connectors, add the Filesystem connector, point it at your vault path. Claude will ask for your approval before it touches a file, every time.

An Obsidian plugin, the embedded way. Community plugins like Claudian or Claude Code IDE run Claude inside Obsidian itself, your vault becomes its working directory and you never leave the app. More moving parts, worth it if you want the chat window next to your notes instead of in a separate terminal.

Three doors, one room. Start with the terminal, add the others if you want them.

05. CLAUDE.md, the brain

Read automatically, every single session

One file in the root of the vault does the actual work: CLAUDE.md. Claude Code reads it automatically at the start of every session in that folder. It tells Claude who you are, how the vault is organised, and the rules to follow, so you are never re-explaining the basics.

# CLAUDE.md

## Who I Am
[Your name, role, main goals]

## Vault Structure
- Inbox = raw captures
- Daily = daily notes
- Projects = active work
- Knowledge = evergreen notes
- Archives = closed out

## Rules
- Start every session by reading today's Daily + Inbox
- Link notes aggressively, do not let anything sit orphaned
- End every session with a /handover summary
- Style: concise, actionable

Keep it short. This file gets read every time, a bloated one wastes tokens on rules Claude already followed the last ten sessions. Update it as your workflow changes, it is a living document, not a one-time setup step.

06. The daily loop

Morning in, evening out

The habit that actually makes this stick is short and repeats every day. Morning: braindump into today's Daily note, whatever is on your mind, unfiltered. Then ask Claude to process the Inbox and Daily, pull out real tasks, and link anything related to existing notes.

Evening: type /handover. Claude writes a summary of what happened and what is next, saved into the vault. Tomorrow's session opens by reading that file first, so it never starts blind.

07. Research ingestion

The flow that actually compounds the vault

Drop a source into the Inbox, an article, a transcript, a link. Then ask Claude to ingest it: pull the key points, and connect them to whatever already exists in Knowledge and Projects. This is the step that turns a pile of notes into an actual network.

A vault that only ever receives new notes stays flat. One that gets re-read and re-linked every time something new comes in starts compounding, the fifth note you add is more useful than the first, because it connects to four others.

The ingestion prompt matters more than people expect. "Summarise this" gives you a summary that sits alone. "Connect this to what I already know" gives you a link, and a link is the whole point of a graph.

08. Five starter skills

Drop these into ~/.claude/skills

A Claude Code skill is a folder with one required file, SKILL.md, a short YAML header plus a workflow in plain English. Claude auto-loads it the moment your prompt matches the description, no slash command to memorise. Save each of these as its own folder under ~/.claude/skills/<name>/SKILL.md and restart Claude Code once to pick them up.

# ~/.claude/skills/weekly-review/SKILL.md
---
name: weekly-review
description: Use when the user asks for a weekly review or to reflect on the past week.
---
Read every Daily note from the last 7 days. Summarise what shipped
against what is still open. Flag any pattern worth noticing. Draft a
short weekly note in 03 Knowledge, linked back to each Daily note used.

# ~/.claude/skills/ingest/SKILL.md
---
name: ingest
description: Use when the user drops a source into Inbox and wants it processed.
---
Read the source in 00 Inbox. Pull out the key points. Search 03
Knowledge for related notes and link both ways. Move the processed
note out of Inbox into Knowledge or Projects, whichever fits.

# ~/.claude/skills/handover/SKILL.md
---
name: handover
description: Use at the end of a session when the user types /handover.
---
Summarise what happened this session in 2 to 3 lines. Write it into
today's Daily note under a Handover heading. List the next concrete
step, so tomorrow's session reads this first and starts already briefed.

# ~/.claude/skills/research/SKILL.md
---
name: research
description: Use when the user wants a topic researched and turned into a vault note.
---
Search for the topic, synthesise findings into a concise note with
sources. Link it to every related note already in 03 Knowledge.
Save into Knowledge, not Inbox, this is a finished note, not a capture.

# ~/.claude/skills/brainstorm/SKILL.md
---
name: brainstorm
description: Use when the user wants new ideas connected to existing vault content.
---
Scan 03 Knowledge and 02 Projects for notes related to the prompt.
Combine existing ideas in ways not yet written down. Draft an ideas
note in Projects, linked back to every source note it pulled from.

/resume. Picks up exactly where the last session left off, using the handover note as its starting context.

Graph view. Once notes are actually linked by the skills above, this stops being a gimmick and starts being a real map of how your ideas connect.

09. Five ways vaults die

Seen in almost every abandoned vault

No CLAUDE.md

Every session starts from zero again. The whole point of the setup, skipped.

One giant folder

No Inbox, no Archives, everything mixed together. Unreadable to you and to Claude both.

Never linking notes

The graph view stays empty. Nothing compounds, it is just a pile of separate files.

Everything on day one

Slash commands, sub-agents, five plugins, before the daily habit even exists. Complexity kills it before it starts.

Skipping the handover

No evening summary means tomorrow's Claude has no idea what happened today. Same as not having a vault at all.

10. Start minimal

Vault, CLAUDE.md, evening handover. That is the whole starting kit, and it is enough to feel the difference in three days. Add one feature a week after that: a slash command when you notice yourself typing the same request, a skill when a task repeats often enough to earn one.

Write notes for yourself first. Do not build the vault to impress a future audit of your own productivity, build it because today's version of you needed to get an idea out of your head. Let Claude organise around what you actually wrote, not the other way round.

There are two kinds of notes now. The ones sitting in a folder, and the ones an AI can actually use.