Build an AI Second Brain with Claude

Build a self-maintaining knowledge base Claude reads and writes for you.

Introduction

30 Second Summary

Every time you chat with AI, you start from scratch. Your previous conversations, decisions, and context disappear the moment the session ends. Imagine if AI could remember everything about your business and get smarter over time.

In this project, you will build a persistent knowledge base on your computer that Claude Cowork reads, writes, and maintains for you. Every source you feed it, every decision you log, and every thought you capture becomes part of a structured, interlinked wiki that grows with every conversation.

What You'll Build

A self-maintaining "second brain" powered by Claude Cowork that stores, cross-references, and updates your business knowledge automatically.

By the end of this project, you'll have:

  • A workspace folder with a three-layer architecture: raw sources, wiki pages, and a schema that defines how Claude maintains your knowledge base.
  • A CLAUDE.md schema file that turns Claude into a disciplined wiki maintainer instead of a generic chatbot.
  • Three custom skills (Capture, Ingest, and Lint) that let you quick-log thoughts, process sources into wiki pages, and health-check your knowledge base.
  • Secret Mission: Build a Weekly Review skill that synthesizes your week's captures into a strategic briefing.

Are there any prerequisites?

You need a Claude Pro, Max, Team, or Enterprise subscription. Claude Pro costs $20/month. If you don't have one yet, the first step walks you through signing up.

Set Up Claude Cowork and Your Workspace

To build a knowledge base that grows smarter over time, you need two things: a tool that can read and write files on your computer, and a folder structure for it to work inside.

In this step, you will install the Claude desktop app, create a workspace folder on your computer, and connect Claude Cowork to that workspace so it has read-write access to everything inside it.

In this step, get ready to:
  • Install the Claude desktop app and sign in with your subscription.
  • Create the workspace folder structure on your Desktop.
  • Connect Claude Cowork to your workspace and confirm it can see your files.
Install the Claude desktop app

The Claude desktop app bundles Claude Chat, Claude Cowork, and Claude Code into a single application. You need it installed on your computer to use Cowork's file access features.

  • Open your browser and go to claude.ai/download.
  • Click Download for macOS or Download for Windows depending on your operating system.
  • Open the downloaded file to install the app.
  • Once installed, open the Claude app. Press Cmd+Space (macOS) or the Windows key (Windows) to open your search bar.
  • Type Claude and press Enter to open it.
  • Sign in with the account that has your Claude Pro (or higher) subscription.

What is Claude Cowork?

Claude Cowork is a mode inside the Claude desktop app that gives Claude access to files on your computer. It can read, write, and organize files in folders you grant access to.

Unlike Claude Chat in the browser (which forgets everything between sessions), Cowork works directly with your local files. This is what makes a persistent knowledge base possible.

Don't have a paid Claude plan?

Claude Cowork requires a Claude Pro ($20/month), Max, Team, or Enterprise subscription. If you only have the free plan, you will need to upgrade before continuing.

How do I upgrade my Claude plan to Pro?

Create the workspace folder structure

Your knowledge base lives inside a single folder called second-brain. Inside it, three subfolders separate raw source material from the wiki Claude maintains, and the skills that define how Claude behaves.

  • Open your terminal. Press Cmd+Space (macOS) or the Windows key (Windows) to open your search bar.
  • Type Terminal (macOS) or PowerShell (Windows) and press Enter to open it.

Why use a terminal for this?

One of the subfolders (.claude/skills/) starts with a dot, which makes it hidden in your file browser by default. Creating it via a terminal command is the simplest way to ensure it exists.

  • Move to your Desktop by running this command:

macOS

cd ~/Desktop

Windows

cd ~\Desktop
  • Create the second-brain folder and all its subfolders by running these commands:

macOS

mkdir -p second-brain/raw second-brain/wiki second-brain/.claude/skills

Windows

mkdir second-brain\raw
mkdir second-brain\wiki
mkdir second-brain\.claude\skills

What do these folders do?

  • raw/ holds source material you drop in. Meeting notes, documents, email threads. Claude never modifies files here.
  • wiki/ is where Claude builds and maintains your structured knowledge base. Claude owns this layer entirely.
  • .claude/skills/ holds your custom skills. These are reusable prompts that tell Claude how to perform specific workflows.
  • Confirm the folders exist by running:

macOS

ls -a second-brain

Windows

ls second-brain -Force

You should see .claude, raw, and wiki listed in the output.

Don't see all three folders?

Make sure you ran the mkdir command while inside the ~/Desktop directory. If you see only raw and wiki but not .claude, check that you used the -a flag (macOS) or -Force flag (Windows) to show hidden items.

Help me troubleshoot my second-brain folder structure

Point Claude Cowork at your workspace

Now that the folder structure exists, you need to tell Claude Cowork to use it as a working directory. This grants Claude read-write access to everything inside second-brain/.

  • Switch back to the Claude desktop app.
  • Start a new Cowork session.
  • When prompted to choose a working directory, select the second-brain folder on your Desktop.

What does selecting a working directory do?

Cowork runs inside a sandboxed virtual machine on your computer. The only files it can see are the ones in the folder you select. Nothing else on your computer is accessible to Claude.

This means your private files stay private. Claude only touches what you explicitly share.

  • Verify Claude can see your workspace by typing this message in the Cowork chat:
What's in this workspace?

Claude should respond by listing the contents of your second-brain folder. You should see it mention raw/, wiki/, and .claude/ (or .claude/skills/) in its answer.

Claude can't see any files?

Make sure you selected the correct folder. The working directory should be the second-brain folder itself, not your Desktop. If Claude reports an empty workspace, try starting a new Cowork session and selecting the folder again.

Help me connect Claude Cowork to my workspace

Your workspace is ready. Next up, you will write the schema that tells Claude how to behave as a knowledge worker.

Write the Schema (CLAUDE.md)

Your workspace is ready, but right now Claude has no idea how to use it. Without rules, Claude would dump information randomly across your folders, forget where things are, and create duplicate pages.

In this step, you will create the schema that turns Claude from a generic chatbot into a disciplined wiki maintainer. The schema is a CLAUDE.md file that Claude reads at the start of every session. It defines the architecture, the page conventions, and the behaviors Claude must follow.

In this step, get ready to:
  • Create the CLAUDE.md schema that defines how Claude maintains the wiki.
  • Create the wiki/index.md master catalog that Claude updates on every ingest.
  • Verify Claude understands and follows the schema.
Create the CLAUDE.md schema file

CLAUDE.md is the most important file in the system. Claude reads it at the start of every session, and it tells Claude exactly how to behave as a knowledge worker.

  • In your Claude Cowork chat, ask Claude to create the schema file by sending this message:
Create a file called CLAUDE.md in the root of this workspace with the following content:
  • Then paste the full schema content below into your message, after the instruction above:
# Second Brain Schema

You are a knowledge worker maintaining a personal wiki for a business owner.
You read, write, and maintain all wiki pages. The user sources material and asks questions. You handle summarization, cross-referencing, filing, and bookkeeping.

## Architecture

This workspace has three layers:

1. **raw/** — Source material the user drops in. Never modify files here.
2. **wiki/** — Structured, interlinked markdown pages you maintain. You own this layer entirely.
3. **CLAUDE.md** (this file) — The schema that defines conventions and workflows.

## Wiki Conventions

### Page types
- **Entity pages** (wiki/entities/): People, companies, products, tools. Filename: kebab-case name.
- **Concept pages** (wiki/concepts/): Ideas, strategies, frameworks, principles.
- **Decision pages** (wiki/decisions/): Specific choices made, with context and rationale.
- **Project pages** (wiki/projects/): Active initiatives with status and stakeholders.
- **Source logs** (wiki/sources/): One per ingested source, summarizing what was extracted.

### Page structure
Every wiki page uses this template:

---
title: [Page Title]
type: [entity|concept|decision|project|source]
created: [YYYY-MM-DD]
updated: [YYYY-MM-DD]
sources: [list of source filenames this page draws from]
---

# [Page Title]

[2-3 sentence summary]

## Details

[Main content organized with headers as needed]

## Connections

- Related: [[page-name]], [[other-page]]
- See also: [[related-concept]]

### Cross-linking rules
- Use [[page-name]] syntax for internal links.
- When you mention an entity or concept that has its own page, always link it.
- When you mention something that SHOULD have a page but doesn't, note it as a gap: [[page-name]]* (asterisk marks missing pages).

### Index (wiki/index.md)
The index is the master catalog. Update it on every ingest. Format:

## Entities
- [[entity-name]] — one-line summary (sources: N)

## Concepts
- [[concept-name]] — one-line summary (sources: N)

[...same for decisions, projects, sources]

## Behaviors

- Never create duplicate pages. If information about an existing entity arrives, UPDATE the existing page.
- When new information contradicts existing information, note the contradiction explicitly with dates and sources.
- Keep summaries concise. The wiki should be scannable.
- Every claim should trace back to a source. Note the source filename in the page's frontmatter.
- When the user asks a question, consult the wiki first (read index.md, then relevant pages) before answering.

What does this schema do?

The schema has three main sections that together turn Claude into a disciplined knowledge worker:

  • Architecture defines the three-layer system: raw/ for source material (read-only), wiki/ for the pages Claude maintains, and CLAUDE.md as the rulebook.
  • Wiki Conventions tells Claude how to name files, structure pages, cross-link with [[page-name]] syntax, and keep the index updated.
  • Behaviors sets ground rules: no duplicates, flag contradictions, trace every claim back to a source, and always consult the wiki before answering questions.
  • After Claude confirms it created the file, open your second-brain/ folder in Finder (macOS) or File Explorer (Windows).
  • Verify that CLAUDE.md exists in the root of the folder (not inside raw/ or wiki/).

Don't see CLAUDE.md in your folder?

  • Files starting with a dot or all-caps may be hidden by default on some systems. On macOS, press Cmd+Shift+. in Finder to reveal hidden files.
  • If Claude created the file inside a subfolder instead of the root, ask Claude to move it: "Move CLAUDE.md to the root of the workspace."
  • If Claude did not create the file at all, make sure your message included both the instruction and the full content. Send it again.

Still stuck? Help me figure out why CLAUDE.md isn't appearing in my second-brain folder.

Why CLAUDE.md instead of a chat message?

You could explain these rules to Claude in every conversation, but you would have to repeat yourself every single session. A CLAUDE.md file is persistent. Claude reads it automatically at the start of every session, so your rules apply forever without you lifting a finger.

Create the wiki index

The index is Claude's table of contents for the entire knowledge base. When Claude needs to find something, it reads index.md first to locate the right page.

  • In your Claude Cowork chat, ask Claude to create the index by sending this message:
Create a file called index.md inside the wiki/ folder with the following content:
  • Paste the full index content below into your message:
---
title: Wiki Index
updated: 2026-06-06
---

# Second Brain Index

This is the master catalog of all wiki pages. Organized by type.

## Entities

_No entities yet. Ingest your first source to populate._

## Concepts

_No concepts yet._

## Decisions

_No decisions yet._

## Projects

_No projects yet._

## Sources

_No sources ingested yet._

---

**Stats**: 0 pages | 0 sources ingested | Last updated: 2026-06-06

What does the index do?

The index is a master catalog organized by page type: Entities, Concepts, Decisions, Projects, and Sources. Every time Claude ingests a new source, it updates this file with links to new pages.

Right now the index is empty because you have not ingested any sources yet. By the end of this project, it will be filled with interlinked references to everything Claude has processed.

  • After Claude confirms it created the file, check your wiki/ folder in Finder or File Explorer.
  • You should see index.md inside the wiki/ folder.

Index.md not showing up in wiki/?

  • Make sure Claude placed the file inside wiki/ and not in the root of your workspace. Ask Claude: "Where did you put the index.md file?"
  • If Claude created a wiki/index/ folder instead of a file, ask it to delete that folder and create wiki/index.md as a file instead.

Need help? Claude didn't put index.md in the right place. Help me fix this.

✔️ Awesome, I've got everything!

Great. Make sure both files are saved. You should have CLAUDE.md in the root and index.md inside wiki/.

ⓧ I'd like to double check the full code

Here is the complete content for both files. Compare yours against these:

CLAUDE.md (in the root of second-brain/):

# Second Brain Schema

You are a knowledge worker maintaining a personal wiki for a business owner.
You read, write, and maintain all wiki pages. The user sources material and asks questions. You handle summarization, cross-referencing, filing, and bookkeeping.

## Architecture

This workspace has three layers:

1. **raw/** — Source material the user drops in. Never modify files here.
2. **wiki/** — Structured, interlinked markdown pages you maintain. You own this layer entirely.
3. **CLAUDE.md** (this file) — The schema that defines conventions and workflows.

## Wiki Conventions

### Page types
- **Entity pages** (wiki/entities/): People, companies, products, tools. Filename: kebab-case name.
- **Concept pages** (wiki/concepts/): Ideas, strategies, frameworks, principles.
- **Decision pages** (wiki/decisions/): Specific choices made, with context and rationale.
- **Project pages** (wiki/projects/): Active initiatives with status and stakeholders.
- **Source logs** (wiki/sources/): One per ingested source, summarizing what was extracted.

### Page structure
Every wiki page uses this template:

---
title: [Page Title]
type: [entity|concept|decision|project|source]
created: [YYYY-MM-DD]
updated: [YYYY-MM-DD]
sources: [list of source filenames this page draws from]
---

# [Page Title]

[2-3 sentence summary]

## Details

[Main content organized with headers as needed]

## Connections

- Related: [[page-name]], [[other-page]]
- See also: [[related-concept]]

### Cross-linking rules
- Use [[page-name]] syntax for internal links.
- When you mention an entity or concept that has its own page, always link it.
- When you mention something that SHOULD have a page but doesn't, note it as a gap: [[page-name]]* (asterisk marks missing pages).

### Index (wiki/index.md)
The index is the master catalog. Update it on every ingest. Format:

## Entities
- [[entity-name]] — one-line summary (sources: N)

## Concepts
- [[concept-name]] — one-line summary (sources: N)

[...same for decisions, projects, sources]

## Behaviors

- Never create duplicate pages. If information about an existing entity arrives, UPDATE the existing page.
- When new information contradicts existing information, note the contradiction explicitly with dates and sources.
- Keep summaries concise. The wiki should be scannable.
- Every claim should trace back to a source. Note the source filename in the page's frontmatter.
- When the user asks a question, consult the wiki first (read index.md, then relevant pages) before answering.

wiki/index.md (inside the wiki/ folder):

---
title: Wiki Index
updated: 2026-06-06
---

# Second Brain Index

This is the master catalog of all wiki pages. Organized by type.

## Entities

_No entities yet. Ingest your first source to populate._

## Concepts

_No concepts yet._

## Decisions

_No decisions yet._

## Projects

_No projects yet._

## Sources

_No sources ingested yet._

---

**Stats**: 0 pages | 0 sources ingested | Last updated: 2026-06-06
Test the schema

The schema only works if Claude actually reads and follows it. Let's confirm Claude loads your CLAUDE.md by asking it about the wiki conventions.

  • In your Claude Cowork chat, start a new message and ask:
What are the rules for maintaining this wiki?
  • Read Claude's response. It should describe your schema conventions back to you.

✔️ Claude described my wiki conventions

Claude should mention things like the three-layer architecture, the page types (entity, concept, decision, project, source), the cross-linking rules, and the behaviors (no duplicates, flag contradictions, trace claims to sources).

If Claude mentions these, your schema is loaded and working. Every future session in this workspace will follow these rules automatically.

ⓧ Claude gave a generic answer

If Claude responds with generic information instead of YOUR specific conventions, it likely did not read the CLAUDE.md file. Try these fixes:

  • Make sure CLAUDE.md is in the root of your second-brain/ folder (not inside a subfolder).
  • Confirm Claude Cowork is still pointed at the second-brain/ folder as its working directory. If you are unsure, ask Claude: "What workspace are you currently working in?"
  • Try starting a new session in Claude Cowork. Claude reads CLAUDE.md at the start of each session, so a fresh session may pick it up.

Need more help? Claude isn't reading my CLAUDE.md file. Help me troubleshoot.

Your schema is locked in. Claude now knows the architecture, the conventions, and the rules. Next up, you will build the Capture skill so you can quickly log thoughts and decisions into the system.

Build the Capture Skill

Your schema is loaded and Claude knows the rules. Now you need a way to quickly log thoughts, decisions, and meeting notes without triggering a full wiki rebuild.

The Capture skill solves this. It gives Claude a repeatable workflow for writing quick entries to raw/captures/ in a consistent format. Every capture gets timestamped, categorized, and tagged with relevant entities so nothing slips through the cracks.

In this step, get ready to:
  • Create the Capture skill file that teaches Claude how to log quick entries.
  • Test the skill with a real capture (a decision, thought, or meeting note).
  • Verify the captured file has the correct structure and location.
Create the Capture skill file

Skills in Claude Cowork are markdown files that live in .claude/skills/. Each skill has a SKILL.md file with a description in the frontmatter and instructions in the body. Claude reads skill descriptions at the start of every session so it knows what's available. The full skill content only loads when you invoke it.

  • In your Claude Cowork session, ask Claude to create the skill file by sending this message:
Create a file at .claude/skills/capture/SKILL.md with this exact content:

---
description: Quick-log a thought, decision, meeting note, or observation into raw captures
---

# Capture

The user wants to quickly log something. Take what they say and write it as a structured capture file.

## Instructions

1. Create the file at `raw/captures/YYYY-MM-DD-[short-slug].md` using today's date and a 2-4 word slug from the content.
2. Use this format:

---
type: [decision|thought|meeting|observation]
date: [YYYY-MM-DD]
entities: [list any people, companies, or projects mentioned]
---

# [Brief title]

[The user's content, cleaned up for clarity but preserving their voice and meaning]

3. After writing, confirm what you captured with a one-line summary.
4. Do NOT ingest the capture into the wiki yet. Captures accumulate in raw/ until the user explicitly runs Ingest.

What does this skill define?

  • The description frontmatter field is what Claude sees at session start. It tells Claude when to use this skill.
  • The instructions tell Claude to write captures to raw/captures/ with a date-stamped filename and structured frontmatter (type, date, entities).
  • The "do NOT ingest" rule keeps captures separate from the wiki until you explicitly process them later.
  • Wait for Claude to confirm it created the file.
  • Ask Claude to show you the contents of .claude/skills/capture/SKILL.md to confirm it matches the content above.

Claude didn't create the file?

If Claude says it cannot create files, make sure your second-brain/ folder is set as the workspace in Claude Cowork. Claude can only write to folders you've explicitly granted access to.

If Claude created the file but the content doesn't match, ask Claude to overwrite it with the exact content you provided. Small differences in formatting can affect how the skill behaves.

Help me troubleshoot why Claude can't create files in my workspace

Test the skill with a real capture

The best way to confirm a skill works is to use it. You'll tell Claude to capture a real decision or thought, and watch it follow the instructions from your new skill.

  • Send Claude a capture request. Use something real to you, or try this example:
Capture: I decided to raise prices 15% on the premium tier starting next month. The reasoning is that our support costs have doubled and we're undercharging relative to competitors.

Why use something real?

Your second brain is most useful when it contains actual information about your business. Starting with a real decision, thought, or meeting note means you're building value from the very first capture.

If you don't have a real example handy, the pricing decision above works perfectly for testing.

  • After sending, Claude should write a file to raw/captures/ and confirm with a one-line summary of what it captured.

Claude didn't follow the capture format?

If Claude wrote the capture but didn't use the defined format (missing frontmatter, wrong file path, etc.), it may not have loaded the skill. Try starting a new session in the same workspace. Skills are discovered at session start.

You can also explicitly invoke the skill by saying: "Use the Capture skill to log this."

Help me get Claude to use my Capture skill correctly

Verify the captured file
  • Ask Claude to show you the file it just created by sending:
Show me the contents of the capture file you just wrote.

You should see a markdown file with three key parts: frontmatter metadata (type, date, entities), a brief title, and your content cleaned up for clarity.

  • Confirm the file is located in raw/captures/ with a filename that follows the YYYY-MM-DD-[short-slug].md pattern.
  • Confirm the frontmatter includes type, date, and entities fields.
  • Confirm your content is preserved (cleaned up for clarity, but your meaning intact).

What should a good capture look like?

If you used the pricing example, your file might be named something like 2026-06-08-raise-prices-premium.md. The frontmatter should have type: decision and list any companies or people you mentioned in the entities field.

The exact filename and date will vary based on when you run this. What matters is the structure: date-stamped name, typed frontmatter, and clear content.

You can also browse your second-brain/ folder in Finder (macOS) or File Explorer (Windows) to see the new raw/captures/ folder and file for yourself.

✔️ Awesome, I've got everything!

Great. Make sure your capture file is sitting in raw/captures/ with the correct format before moving on.

ⓧ I'd like to double check the full code

Here's what your SKILL.md file should contain:

---
description: Quick-log a thought, decision, meeting note, or observation into raw captures
---

# Capture

The user wants to quickly log something. Take what they say and write it as a structured capture file.

## Instructions

1. Create the file at `raw/captures/YYYY-MM-DD-[short-slug].md` using today's date and a 2-4 word slug from the content.
2. Use this format:

---
type: [decision|thought|meeting|observation]
date: [YYYY-MM-DD]
entities: [list any people, companies, or projects mentioned]
---

# [Brief title]

[The user's content, cleaned up for clarity but preserving their voice and meaning]

3. After writing, confirm what you captured with a one-line summary.
4. Do NOT ingest the capture into the wiki yet. Captures accumulate in raw/ until the user explicitly runs Ingest.

Your Capture skill is working. You now have a quick way to log anything into your second brain without worrying about organization. Next up, you'll build the Ingest skill that processes these raw captures into structured, cross-linked wiki pages.

Build the Ingest Skill

Your Capture skill is working. You can log thoughts and decisions into raw/captures/ with a single sentence. But those captures are just sitting there, unprocessed.

The Ingest skill is what turns raw material into structured knowledge. When you invoke it, Claude reads a source file, extracts the important entities, concepts, and decisions, then builds (or updates) wiki pages with full cross-references. This is the engine that makes the wiki grow.

In this step, get ready to:
  • Create the Ingest skill that processes sources into wiki pages.
  • Feed your first real source and watch Claude build the wiki.
  • Explore the cross-linked pages Claude created.
Create the Ingest skill file

The Ingest skill is the most complex of the three skills you'll build. It handles reading source files, extracting key information, creating or updating wiki pages, adding cross-links between pages, and keeping the index up to date. All of those behaviors are defined in a single SKILL.md file.

  • Inside your second-brain/.claude/skills/ folder, create a new folder called ingest.
  • Inside the ingest folder, create a new file called SKILL.md.
  • Open SKILL.md in any text editor and paste in the following content:
---
description: Process a raw source and integrate it into the wiki with cross-links and index updates
---

# Ingest

The user wants to process a source from raw/ into the wiki.

## Instructions

1. Read the source file(s) the user specifies (or all unprocessed files in raw/ if they say "ingest all").
2. For each source:
   a. Extract key entities (people, companies, products, tools).
   b. Extract key concepts (strategies, ideas, frameworks).
   c. Extract decisions (choices made with rationale).
   d. Extract project references (initiatives, timelines, goals).
3. For each extracted item:
   - If a wiki page already exists: UPDATE it with the new information. Add the new source to its frontmatter. Note any contradictions.
   - If no wiki page exists: CREATE one using the page template from CLAUDE.md.
4. Add cross-links ([[page-name]]) between all related pages.
5. Create a source log at `wiki/sources/[source-filename].md` summarizing what was extracted.
6. Update `wiki/index.md` with any new pages and updated source counts.
7. Report back: "Ingested [source]. Created N new pages, updated M existing pages. Key entities: [list]."

## Processing captures

When ingesting from raw/captures/:
- Process each capture file as a mini-source.
- Decisions get their own decision pages.
- Entities mentioned get added to or linked from entity pages.
- Thoughts and observations enrich existing concept pages or create new ones if the topic is substantial enough.
- After ingesting, note which capture files were processed (but do not delete them from raw/).

What does this skill do?

The Ingest skill has two main sections:

  • Instructions defines a 7-step process: read the source, extract entities/concepts/decisions/projects, create or update wiki pages, add cross-links, create a source log, and update the index.
  • Processing captures gives special rules for ingesting from raw/captures/. Decisions get their own pages. Thoughts enrich concept pages. Nothing in raw/ is ever deleted.
  • Save SKILL.md.
  • Confirm the file is in the right place by opening second-brain/.claude/skills/ingest/ in Finder (macOS) or File Explorer (Windows). You should see SKILL.md listed alongside your capture folder.

Don't see the ingest folder?

  • Make sure you created the ingest folder inside .claude/skills/, not inside the capture folder or the root of second-brain.
  • The .claude folder is hidden by default on most systems. On macOS, press Cmd+Shift+. in Finder to show hidden files. On Windows, click View and check Hidden items.

Help me find where my SKILL.md file should go in the second-brain folder structure.

Ask Claude to create it for you

Since Claude Cowork has write access to your workspace, you can also paste the skill content into the chat and ask Claude to save it as .claude/skills/ingest/SKILL.md. Claude will create the folder and file for you.

✔️ Awesome, I've got everything!

Great. Double check you've saved your file before moving on.

ⓧ I'd like to double check the full code

Here's the complete SKILL.md file for your Ingest skill:

---
description: Process a raw source and integrate it into the wiki with cross-links and index updates
---

# Ingest

The user wants to process a source from raw/ into the wiki.

## Instructions

1. Read the source file(s) the user specifies (or all unprocessed files in raw/ if they say "ingest all").
2. For each source:
   a. Extract key entities (people, companies, products, tools).
   b. Extract key concepts (strategies, ideas, frameworks).
   c. Extract decisions (choices made with rationale).
   d. Extract project references (initiatives, timelines, goals).
3. For each extracted item:
   - If a wiki page already exists: UPDATE it with the new information. Add the new source to its frontmatter. Note any contradictions.
   - If no wiki page exists: CREATE one using the page template from CLAUDE.md.
4. Add cross-links ([[page-name]]) between all related pages.
5. Create a source log at `wiki/sources/[source-filename].md` summarizing what was extracted.
6. Update `wiki/index.md` with any new pages and updated source counts.
7. Report back: "Ingested [source]. Created N new pages, updated M existing pages. Key entities: [list]."

## Processing captures

When ingesting from raw/captures/:
- Process each capture file as a mini-source.
- Decisions get their own decision pages.
- Entities mentioned get added to or linked from entity pages.
- Thoughts and observations enrich existing concept pages or create new ones if the topic is substantial enough.
- After ingesting, note which capture files were processed (but do not delete them from raw/).
Feed your first real source

The Ingest skill needs something to process. You'll drop a real source document into raw/ and then ask Claude to ingest it.

  • Pick a real source document from your work. This could be a business plan, meeting notes, a competitor analysis, a strategy document, or even a pasted email thread.
  • Save or copy that source file into your second-brain/raw/ folder. If it's text, you can create a .md or .txt file and paste the content in.
  • Switch back to Claude Cowork and tell Claude to ingest your source. For example, type something like: "Ingest the file raw/my-business-plan.md" (using your actual filename).

What should I see?

Claude will read your source, then start creating files in the wiki/ folder. You'll see Claude report back something like: "Ingested my-business-plan.md. Created 4 new pages, updated 0 existing pages. Key entities: [list of names and companies it found]."

Claude creates entity pages in wiki/entities/, concept pages in wiki/concepts/, and a source log in wiki/sources/. It also updates wiki/index.md with all the new pages.

Claude didn't create any wiki pages?

  • Check that your source file is actually inside the second-brain/raw/ folder. Claude can only read files within the workspace.
  • Make sure the filename you told Claude matches the actual file. Try asking "what files are in raw/" to see what Claude can see.
  • If your source is very short (a sentence or two), Claude might not find enough to extract. Try a more substantial document with names, companies, or decisions mentioned.

Claude isn't creating wiki pages when I ask it to ingest my source file. Help me troubleshoot.

Explore what Claude built

Your wiki now has real content. Time to see how the pieces connect.

  • Ask Claude: "Show me the index."
  • Claude will display the updated wiki/index.md, which now lists every page it created along with one-line summaries and source counts.
  • Open your second-brain/wiki/ folder in Finder (macOS) or File Explorer (Windows).
  • Browse into wiki/entities/ and open one of the entity pages in a text editor. You should see frontmatter (title, type, created date, sources), a summary, details, and a Connections section with [[page-name]] links to related pages.
  • Check wiki/sources/ for the source log. This file summarizes everything Claude extracted from your original document.

Why cross-links matter

The [[page-name]] links between pages are what make this a knowledge graph instead of a flat folder of notes. When Claude ingests new sources later, it follows these links to find related pages and update them. The more connected the wiki gets, the smarter Claude's responses become.

Your wiki is alive. Claude just turned a raw document into structured, interlinked knowledge pages. Next, you'll build the Lint skill to keep your growing wiki healthy and contradiction-free.

Build the Lint Skill

Your Capture and Ingest skills give Claude the ability to add knowledge to the wiki. But how do you know when the wiki needs attention?

A knowledge base without maintenance slowly decays. Pages become orphaned, information contradicts itself, and gaps go unnoticed. The Lint skill solves this by giving Claude the ability to audit the wiki and surface problems before they compound.

In this step, get ready to:
  • Create the Lint skill that health-checks your wiki for contradictions, orphans, and gaps.
  • Run your first lint and review Claude's findings.
  • Act on a lint suggestion to demonstrate the compounding loop.
Create the Lint skill

The Lint skill is your wiki's quality control system. When invoked, Claude scans every page for six categories of issues: orphans, contradictions, stale content, missing cross-references, gaps, and weak pages.

  • In Finder (macOS) or File Explorer (Windows), navigate to the .claude/skills/ folder inside your second-brain directory.
  • Create a new folder called lint inside .claude/skills/.
  • Inside the lint folder, create a new file called SKILL.md.
  • Open SKILL.md in any text editor and paste the following content:
---
description: Run health checks on the wiki to find contradictions, orphans, gaps, and stale pages
---

# Lint

The user wants a health check on the wiki.

## Instructions

1. Read `wiki/index.md` to get the full page inventory.
2. Read each wiki page and check for:

### Orphan pages
Pages with no inbound links from other pages. These are isolated and may need connecting.

### Contradictions
Places where two pages make conflicting claims. Flag with the specific quotes and sources.

### Stale pages
Pages whose `updated` date is more than 30 days old relative to today. These may need refreshing.

### Missing links
Entity or concept names mentioned in page text that have their own wiki page but are not linked with [[brackets]].

### Gaps
Topics referenced with the * marker ([[page-name]]*) or entities mentioned repeatedly across pages that don't have their own page yet.

### Weak pages
Pages with very little content (under 3 sentences in Details) that could be enriched.

3. Report findings grouped by category. For each finding, suggest a specific action.
4. Ask the user which findings they'd like to address now.

## Tone

Be direct and useful. Think of yourself as an editor reviewing a knowledge base, not a teacher giving feedback. "The page for [X] contradicts the page for [Y] on pricing — [X] says $50/month, [Y] says $45/month. Which is current?"

What does this skill do?

The Lint skill tells Claude to audit your entire wiki when you ask for a health check. It looks for six specific categories of issues:

  • Orphan pages have no inbound links from other pages, meaning they're isolated from the rest of the wiki.
  • Contradictions are places where two pages make conflicting claims about the same thing.
  • Gaps are topics that get mentioned across pages but don't have their own dedicated page yet.

The Tone section at the bottom shapes how Claude communicates findings. It aims for direct, editor-style feedback with specific quotes rather than vague summaries.

  • Save SKILL.md and close the text editor.
  • Confirm the file is in the right place. Your .claude/skills/ folder should now contain three subfolders: capture, ingest, and lint.

Don't see all three skill folders?

Make sure the lint folder is inside .claude/skills/ (not at the root of second-brain). The path should be second-brain/.claude/skills/lint/SKILL.md.

On macOS, the .claude folder may be hidden because it starts with a dot. Press Cmd+Shift+. in Finder to toggle hidden files.

Help me find my .claude/skills folder in Finder

✔️ Awesome, I've got everything!

Great. Double check you've saved the file, then move on to test it.

ⓧ I'd like to double check the full code

Here's the complete content for .claude/skills/lint/SKILL.md:

---
description: Run health checks on the wiki to find contradictions, orphans, gaps, and stale pages
---

# Lint

The user wants a health check on the wiki.

## Instructions

1. Read `wiki/index.md` to get the full page inventory.
2. Read each wiki page and check for:

### Orphan pages
Pages with no inbound links from other pages. These are isolated and may need connecting.

### Contradictions
Places where two pages make conflicting claims. Flag with the specific quotes and sources.

### Stale pages
Pages whose `updated` date is more than 30 days old relative to today. These may need refreshing.

### Missing links
Entity or concept names mentioned in page text that have their own wiki page but are not linked with [[brackets]].

### Gaps
Topics referenced with the * marker ([[page-name]]*) or entities mentioned repeatedly across pages that don't have their own page yet.

### Weak pages
Pages with very little content (under 3 sentences in Details) that could be enriched.

3. Report findings grouped by category. For each finding, suggest a specific action.
4. Ask the user which findings they'd like to address now.

## Tone

Be direct and useful. Think of yourself as an editor reviewing a knowledge base, not a teacher giving feedback. "The page for [X] contradicts the page for [Y] on pricing — [X] says $50/month, [Y] says $45/month. Which is current?"
Run your first lint

Your wiki currently has pages from just one ingested source. That makes this the perfect time to lint. Claude will identify gaps and weak spots while the wiki is still small and manageable.

  • Switch back to Claude Cowork with your second-brain workspace open.
  • Type a message asking Claude to lint the wiki. Something like: "Lint the wiki" or "Run a health check on my knowledge base."
  • Wait for Claude to read through all your wiki pages and report its findings.

Claude will respond with findings grouped by category. Since you only have one ingested source, expect to see mostly gaps and weak pages rather than contradictions.

What should I see?

Claude's lint report should look something like this: a categorized list of findings with specific suggestions. For example, Claude might tell you that a page mentions a concept that has no dedicated page yet, or that an entity page has very little content and could be enriched with more detail.

At the end of the report, Claude will ask which findings you'd like to address. That's your cue for the next substep.

Claude didn't find anything?

If Claude reports zero findings, your wiki might be very small (only one or two pages). Try asking Claude to look specifically for gaps: "Are there any entities or concepts mentioned in page text that don't have their own page yet?"

If Claude doesn't seem to know about the Lint skill at all, confirm your SKILL.md file is saved at the correct path: .claude/skills/lint/SKILL.md.

Help me get my Lint skill working in Claude Cowork

Act on a lint suggestion

This is where the compounding loop clicks into place. Lint identifies a gap, you fill it with Capture or by asking Claude to create a stub page, and the wiki grows richer. The cycle is: capture, ingest, lint, capture more.

  • Read through Claude's lint findings and pick one suggestion that resonates with you.
  • Choose how to address it. You have two options:
  • If you used Capture, follow up by telling Claude to ingest the new capture into the wiki.
  • Verify that the wiki grew. Ask Claude: "Show me the updated index."
  • You should see new entries in wiki/index.md that weren't there before your lint session.

Why does this loop matter?

Traditional wikis die because maintenance is boring and relentless. Nobody wants to check cross-references manually. With the Lint skill, Claude does the tedious auditing work and surfaces exactly what needs your attention.

Each time you address a lint finding, the wiki gets denser. Denser wikis produce richer lint findings. This is the compounding effect that makes the system grow more valuable over time.

Claude created a duplicate page?

If Claude created a new page for something that already has one, remind it of the schema rule: "Check CLAUDE.md. The wiki conventions say never create duplicate pages. If information about an existing entity arrives, update the existing page instead."

Help me fix a duplicate page in my wiki

You now have all three skills working together. Capture logs raw thoughts, Ingest builds structured wiki pages, and Lint keeps everything healthy. Next up, you'll feed the system more sources and watch the wiki compound in real time.

Feed the System and Watch It Grow

Your Lint skill just identified gaps and weaknesses in the wiki. You acted on at least one suggestion, proving the feedback loop works. Now it's time to push the system harder.

In this step, you'll feed Claude multiple sources of different types, log quick thoughts throughout a work session, and then lint again to see how the wiki compounds. This is where the LLM Wiki pattern proves itself. Every new source doesn't just add pages. It enriches existing pages, creates new cross-links, and makes the whole system smarter.

In this step, get ready to:
  • Ingest 2-3 additional sources of different types and watch Claude update existing pages rather than duplicate them.
  • Log quick captures throughout a work session and ingest them into the wiki.
  • Run a second lint to see the compounding effect in action.
Add more sources and ingest them

A knowledge base built from a single source is just a summary. The real value emerges when Claude can connect information across multiple, different types of sources.

  • Gather 2-3 real documents from your work. Aim for different types. Good options include: meeting notes, a team roster or org chart, a strategy document, a competitor analysis, customer research, or a project brief.
  • Drop your first new source into the raw/ folder inside your second-brain/ workspace.
  • Return to Claude Cowork and tell it to ingest the new source. For example, type something like:
Ingest the file raw/team-meeting-notes-june.md

What should I see?

Claude should report back with something like: "Ingested team-meeting-notes-june.md. Created 3 new pages, updated 2 existing pages. Key entities: [names from your source]." The key word is updated. If an entity from your first source appears again, Claude adds to the existing page instead of creating a duplicate.

  • Drop your second source into raw/ and tell Claude to ingest it.
  • After each ingest, ask Claude to show you the index by typing:
Show me the current wiki index
  • Notice how the source counts increase on existing pages. A page that previously said "(sources: 1)" should now say "(sources: 2)" if that entity appeared in both sources.
  • Drop your third source (if you have one) into raw/ and ingest it the same way.
  • Browse a wiki page that was updated by multiple sources. Ask Claude to show you one by typing:
Show me the wiki page that has the most sources

You should see a page with multiple entries in its sources frontmatter, a richer Details section, and cross-links to pages that didn't exist after your first ingest.

Why does updating matter more than creating?

Most note-taking tools create a new note every time you save something. Over time, you end up with five different notes about the same topic, scattered across dates and folders.

Your CLAUDE.md schema explicitly tells Claude to update existing pages instead of duplicating. This means each entity and concept has one canonical page that gets richer over time.

Claude created duplicate pages?

If Claude created a new page for something that already had one, it may have used a different name (e.g., "acme-corp" vs "acme-corporation"). Ask Claude to merge the duplicates and pick one canonical name. The Lint skill catches these too.

Help me figure out why Claude made a duplicate page instead of updating the existing one.

Use Capture throughout a work session

Formal documents are only half the picture. The other half is the thinking that happens between documents: quick decisions, observations, ideas you don't want to forget.

  • Over the next few minutes, log 3-4 quick captures as if you were working through your day. Use the Capture skill each time by telling Claude something real. For example:
Capture: Had a call with the design team. They want to push the rebrand launch from July to September because the logo explorations aren't ready. I agreed — rushing the brand feels riskier than delaying.
  • Log a second capture. Try a different type (a thought or observation):
Capture: Noticed that our highest-value customers all came from the webinar funnel, not paid ads. Might be worth doubling down on educational content instead of increasing ad spend.
  • Log a third and fourth capture. These could be meeting takeaways, a decision about your product, an observation about your market, or anything relevant to your work.
  • After logging all your captures, verify they landed in the right place by asking Claude:
What captures are in raw/captures/ that haven't been ingested yet?

Claude should list your new capture files. Each one should have today's date and a short descriptive slug in the filename.

  • Now ingest all the captures into the wiki by telling Claude:
Ingest all unprocessed captures

What happens when captures get ingested?

Claude processes each capture as a mini-source. Decisions become their own decision pages. Entities mentioned get linked to existing entity pages (or create new ones). Thoughts and observations enrich concept pages.

This is the magic of the system. A 30-second capture during a phone call becomes structured, linked knowledge that Claude can reference in future conversations.

  • Ask Claude to show you a wiki page that was enriched by your captures. Look for pages where your informal thinking now sits alongside formal document content.

Captures didn't create any new pages?

If your captures only updated existing pages, that's actually a good sign. It means Claude recognized the entities and concepts already in your wiki. Check that the existing pages now have your capture content woven in. If captures seem to have been ignored entirely, ask Claude to show you the source log for one of the capture files.

Why didn't my captures create new wiki pages after ingestion?

Run lint again and compare to your first lint

Your first lint ran against a wiki built from a single source. Now you've ingested multiple sources and a batch of captures. The lint results should tell a completely different story.

  • Tell Claude to run the Lint skill again:
Lint the wiki
  • Compare these results to your first lint. Look for three specific changes:
  • Fewer orphan pages. Pages that were isolated before should now have inbound links from the new sources.
  • More specific gaps. Instead of broad suggestions like "you need more pages," Claude should identify precise missing topics based on patterns across your sources.
  • Possible contradictions. If two sources say different things about the same entity, Claude should flag the conflict with specific quotes.

What does a healthy lint look like?

A perfect lint (zero findings) isn't the goal. A useful lint identifies the next highest-value actions for your knowledge base. As the wiki grows, lint findings shift from "you're missing basic pages" to "these two pages contradict each other on pricing" or "this project hasn't been updated in 30 days."

  • Pick one lint finding and address it. You can either capture new information to fill a gap, ask Claude to fix a missing cross-link, or resolve a contradiction by telling Claude which version is correct.

You've now completed the full compounding loop: capture, ingest, lint, capture more. Each cycle makes the wiki richer and Claude's understanding of your work deeper.

Why does this compound?

Every time you open a new Claude Cowork session in this workspace, Claude reads your CLAUDE.md and checks wiki/index.md to understand your full knowledge base. It never starts from zero. Tomorrow's session builds on today's wiki. Next week builds on this week.

Secret mission

Build a Weekly Review Skill

Your wiki captures daily knowledge, but who connects the dots across a whole week? Build a Weekly Review skill that synthesizes recent activity into a strategic briefing. Think chief of staff summarizing the week for a CEO.

Clean Up Your Resources

Clean Up Your Resources

This project runs entirely on your local machine. Your Claude Pro subscription is an existing cost you already had before starting. There are no additional ongoing charges from the files you created.

Decide whether to keep your knowledge base growing, pause your Cowork session, or delete everything entirely.

Resources you used:

  • second-brain/ folder on your Desktop (contains all wiki pages, skills, captures, raw sources, and the CLAUDE.md schema).

Keep everything running

No action needed. Choose this if you want your knowledge base to keep compounding.

Every time you open Claude Cowork and point it at your second-brain/ folder, it picks up right where you left off. The Claude desktop app stays installed for your other work too.

  • Keep feeding sources into raw/ and running Ingest to grow your wiki.
  • Use Capture throughout your day to log decisions and observations.
  • Run Lint weekly to keep your knowledge base healthy.
  • Run Weekly Review every Friday to synthesize patterns.

Pause - I'll come back to this later

Close your Claude Cowork session to free up system resources. Your files stay on disk and Claude Cowork can reconnect to them any time.

  • Close the Claude desktop app (or just close the Cowork session).
  • Your second-brain/ folder remains on your Desktop with all wiki pages, skills, and sources intact.
  • To resume later, open Claude Cowork and select your second-brain/ folder as the working directory again.

Delete - I don't want to use this again

Remove all project files from your machine. This permanently deletes your knowledge base, skills, and all wiki content.

  • Close the Claude desktop app.
  • Delete the second-brain/ folder by running this command in your terminal:

macOS

rm -rf ~/Desktop/second-brain

Windows

Remove-Item -Path "$env:USERPROFILE\Desktop\second-brain" -Recurse -Force

What about the Claude desktop app?

You likely want to keep the Claude desktop app installed for other uses. It does not store any project-specific data outside your second-brain/ folder, so deleting the folder is a complete cleanup.

If you do want to remove the app: on macOS, drag Claude from your Applications folder to the Trash. On Windows, open Settings, select Apps, find Claude, and click Uninstall.

Nice Work!

Nice Work!

Nice work! You've built a persistent, self-maintaining second brain powered by Claude Cowork. Every conversation you have from now on builds on everything that came before.

You've learned how to:

  • Build a three-layer knowledge base (raw sources, wiki, schema) with a CLAUDE.md that turns Claude into a disciplined wiki maintainer instead of a generic chatbot.
  • Create Capture and Ingest skills that quick-log decisions and process sources into structured, cross-linked wiki pages.
  • Create a Lint skill that health-checks your knowledge base for contradictions, orphan pages, and knowledge gaps.
  • Build a Weekly Review skill that synthesizes your week's captures and ingests into a strategic briefing.

Ready to quiz yourself?