🧠 How to Build a Second Brain with Claude

What a Second Brain with Claude Actually Means

Most people use Claude the same way they use a search engine. Ask a question, get an answer, close the tab. The conversation is fresh every time. Claude has no idea what you worked on yesterday, what decisions you made last month, or what kind of outputs you actually like.

A second brain changes this. It gives Claude structured access to your real context so that instead of answering generically, it reasons from your actual files, priorities, and knowledge. The difference between those two things is the difference between a smart stranger and a well-briefed partner.

Disclaimer: This setup assumes Claude is connected to your tools (like Google Drive or Notion) through MCP or a similar integration, so it can access your files during a session.

For example, I asked Claude to analyze my existing content materials, using the files I provided, and turn everything into a clear, prioritized weekly action plan:

The architecture has three layers. They're the same regardless of which tools you use.

One Brain or Many

Most people should not build one giant brain for everything. The bigger and more mixed the brain, the worse the retrieval. Claude doesn't know which part of the mess is relevant to your current task unless you tell it exactly where to look.

Build separate brains when the files, decisions, and terminology are different enough to require different retrieval logic.

Pick your first brain before you start. Name it after a role, project, client, or workflow. That name will anchor everything downstream: folder names, instruction files, skill names, and trigger phrases.

Layer 1: The File System

This is where your raw material lives. Documents, transcripts, drafts, meeting notes, research, briefs, templates. Claude reads these directly, so how you organize them matters.

The core principle: Claude retrieves faster and more accurately when your folder structure is described explicitly in your instructions. A chaotic or mixed folder doesn't make retrieval impossible. It makes it unreliable.

File system setup rules:

Create a dedicated top-level folder for the brain. Name it clearly. Inside it, organize by purpose, not by date or project number. These are the subfolders that consistently work across different workflows:

/[brain-name]/
  /active          → current work, drafts, in-progress documents
  /reference       → source material, research, documentation
  /templates       → reusable formats, frameworks, structures
  /archive         → completed work, old versions, closed projects
  /inbox           → unprocessed notes, quick captures, things to sort

The inbox folder deserves specific attention. This is where you drop anything quickly: a voice note transcript, a clipped article, a rough idea. You don't need to organize it in the moment. The whole point is that Claude can process it later in batch.

What to move in on day one:

Don't migrate everything. This is the trap most people fall into and why their system dies in week two. Start with the top 10 to 15 documents that you reference most often in your work. Active project briefs. Client background documents. Your most-used templates. Anything you currently re-explain to Claude manually every session.

Everything else can stay where it is. Expand the brain over time as you notice what's missing.

For example,  I asked Claude to find the three biggest missed opportunities in my content system — and here’s what I got.

Layer 2: The Context System

This is the layer most people skip. It's also the layer that determines whether the brain actually changes how you work or just becomes another folder system.

The context layer stores structured meaning, not raw files. Current priorities. Active tasks. Campaign status. Decisions made. What's in progress, what's blocked, what's done. This is the information Claude needs to prioritize and sequence, not just retrieve.

What belongs in the context layer:

The minimum viable context layer for most knowledge workers has four components:

1. Current priorities A simple ranked list of what matters most right now. Updated weekly. This doesn't need to be elaborate. A plain text file with five to ten items is enough. When Claude reads this before a task, the quality of its recommendations immediately improves.

2. Active projects tracker For each live project: what it is, its current status, the next action, any blockers, and a link to the relevant folder. This can live in Notion, Airtable, a markdown table, or a simple text file. The format matters less than keeping it current.

3. Decisions log A running record of significant decisions and the reasoning behind them. This is the context that's hardest to rebuild once lost. When Claude can see why you made past decisions, it stops recommending things you already tried and rejected.

4. Standing instructions How you like to work. Preferred output formats. Things you never want Claude to do. Tone preferences. Meeting prep format. Report structure. This is the file that replaces the first paragraph of every prompt you've ever written.

Layer 3: The CLAUDE.md File

This is the instruction layer. It's a single markdown file that tells Claude what this brain is for, where to look, when to act, and how to respond. In supported setups (like Claude Code or MCP), Claude reads this file at the start of each session

Getting this file right is the highest-ROI hour you'll spend on this entire system. A weak CLAUDE.md produces generic answers. A well-written one produces outputs you'd actually send.

What the file should contain:

Name and scope What the brain is for and, critically, what it is not for. Both matter. Defining scope explicitly prevents Claude from pulling from the wrong sources or treating adjacent topics as in-scope.

Knowledge architecture Where the files live and what they contain. Where the context lives and what it tracks. Name both systems explicitly. Don't say "my Google Drive." Say "the /content-brain/active/ folder in Google Drive, which contains current drafts and campaign briefs."

Trigger logic What types of requests should activate deep retrieval versus quick answers. What phrases or topics mean "go check the files" versus "just reason from context."

Operating modes This section changes how Claude behaves based on what you're asking. The most useful modes for a knowledge worker are:

• Decision mode: What should I do / choose / prioritize? Output: recommendation, reasoning, tradeoffs, next step.

• Planning mode: Build me a weekly plan / roadmap / execution sequence. Output: priorities in order, quick wins, blockers, plan.

• Creation mode: Draft / write / build something. Output: ready-to-use content, linked source reasoning, what's missing.

• Synthesis mode: Make sense of these sources / find the pattern. Output: what was found, key patterns, strategic takeaway, open questions.

Transparency rules Tell Claude to distinguish between facts from your files, inferences it's drawing, and recommendations it's making. Tell it to say when information is missing or stale. This single instruction prevents the most common failure: Claude confidently presenting guesses as facts from your sources.

A starter CLAUDE.md you can adapt:

# [Brain Name]

You are a high-trust working partner for [scope — your role, project, or workflow].

Your job is to retrieve, organize, reason over, and produce actionable outputs 
from my actual knowledge base — not from generic best practices.

## Knowledge Architecture

**Files:** [your file system name] — raw material, drafts, research, references.
Look in [specific folder paths] for [specific types of content].

**Context:** [your context system] — current priorities, tasks, decisions, status.
Check [specific pages/tables/files] for current state before any planning task.

**Rule:** Use files for substance. Use context for prioritization.
Combine both when the task requires reasoning across current work and source material.
Always retrieve before answering when the request depends on stored knowledge.
If information is missing, stale, or conflicting, say so clearly. Do not fill gaps 
with assumptions presented as facts.

## Operating Modes

**Decision mode** — when I ask what to do, choose, or prioritize:
Give me your best recommendation, why, the tradeoffs, and the next step.

**Planning mode** — when I ask for a plan, roadmap, or execution order:
Give me top priorities in order, quick wins, blockers, and a concrete plan.

**Creation mode** — when I ask for a draft, document, or structured output:
Give me ready-to-use content with source reasoning and flag what's missing.

**Synthesis mode** — when I ask to analyze or make sense of multiple sources:
Give me what you found, the key patterns, the strategic takeaway, and open questions.

## What Not To Do

- Never present inferences as facts from my files
- Never recommend something I've already logged as rejected in my decisions file
- Never pull from [out-of-scope folders or topics]
- Don't ask clarifying questions for straightforward retrieval tasksClaude Command Limitations

The scope section is the most important part. "Never pull from [out-of-scope topics]" does more work than any other line in the file. Without it, a brain that's supposed to be about client work starts pulling from your personal projects. Define the edges.

The Skill Files Worth Building First

The CLAUDE.md defines how the brain behaves overall. Skill files define how Claude handles specific recurring tasks within it. One skill per workflow. Written once. In supported environments, these files can be loaded automatically when relevant.

One well-written skill file is worth more than 50 saved prompts.

These three have the highest return on investment for most knowledge workers. Build them in this order.

Skill 1 — Your Writing Voice

Everything about how you write. Tone, sentence length, structure, words you avoid, what good output looks like. Once this file exists, you never explain your style to Claude again. Every piece of content you produce after building it benefits from it immediately.

What to include:

• Your tone (direct? conversational? formal? blunt?)

• Sentence and paragraph length preferences

• Words or phrases you actively avoid

• What you never want Claude to do in your writing (bullet points everywhere, hedging every claim, starting with "Certainly")

• Two to three examples of writing you're proud of, pasted in full

The examples matter more than the descriptions. Claude will reverse-engineer your style from them.

How to create it (10 minutes):

1. Open any text editor

2. Create a new file: writing-voice.md

3. Write your instructions using the points above — be directive, not conversational

4. Paste in two or three examples of your writing

5. Drop it in your Cowork folder or ~/.claude/skills/

6. Next session, Claude has it

That's the whole process. No complex setup beyond placing the file in the correct folder for your Claude environment.

Skill 2 — Your Recurring Document Format

If you produce any document more than twice a month — reports, client updates, project briefs, strategy docs — this skill locks in its structure permanently. Claude uses it automatically every time you ask for that document type.

What to include:

• Your standard headings in order

• How long each section should be

• What belongs in the opening section

• How you handle recommendations versus findings

• A finished example of a document you're happy with

Skill 3 — Your Weekly Review

This one changes daily work more visibly than the others. A skill that tells Claude how to run your weekly review — which files to check, which context layers to pull, what to surface, what questions to answer — turns a task you skip half the time into a consistent 20-minute session that actually informs the following week.

What to include:

• Which folders to scan for updates

• Which context files to read (priorities, tasks, decisions)

• What questions the review should answer every week

• The format you want back and in what order

• How to flag things that look stale or off-track

Pro tip: Use the /skill-creator command in Cowork if you want Claude to build any of these files with you interactively. It asks questions, writes a draft, and iterates until it's right. Faster than writing from scratch for most people.

The Prompts That Actually Unlock the System

The architecture is the setup. The prompts are where the system earns its keep. These are the ones that show you whether it's working — and the ones worth using daily once it is.

Run these first to test your setup:

These diagnostic prompts tell you whether Claude is reaching your files and context correctly. If the answers feel generic, go back to your CLAUDE.md and make the source names and folder paths more explicit.

• "Using my second brain, tell me the three biggest missed opportunities hidden in my current work."

• "Read my active projects and priorities. What's likely to fall through the cracks this week?"

• "Look at my files and context and tell me what I'm overcomplicating."

Daily use:

• "Using my second brain, build me the smartest weekly action plan you can from what's actually in motion."

• "I need to write [document type]. Check my templates and past examples, then draft it in my voice."

• "Based on my decisions log and current priorities, what should I stop doing?"

• "I have 90 minutes this afternoon. What's the highest-leverage thing I could work on?"

The thinking partner prompt:

This one is different from the others. It's not retrieval. It's using Claude as a reasoning partner before you've figured out what you think.

"I'm working through [problem or decision]. I don't need you to write anything yet. Ask me the questions that would help me think this through more clearly, using what you know about my context and priorities."

Telling Claude explicitly that you're in thinking mode, not writing mode, changes the quality of the conversation significantly. Claude stops producing outputs and starts asking the questions that surface your own thinking. Use this for anything consequential.

Rollout Checklist

Don't try to build the whole system in one afternoon. This is exactly how it dies.

Week 1 — foundations only

• Choose one scope: one role, project, client, or workflow

• Create the folder structure, move in your top 10 to 15 active files

• Write your CLAUDE.md using the starter template above

• Run the three diagnostic prompts and fix what feels off

Week 2 — context layer

• Build your current priorities file

• Build your active projects tracker

• Start your decisions log (even just the last 30 days)

• Update your CLAUDE.md to reference all three by name

Week 3 — skill files

• Write your writing voice skill first

• Write your primary recurring document skill

• Write your weekly review skill

• Run a full weekly review using the complete system

Ongoing

• Update priorities every Monday

• Add to the decisions log when significant calls are made

• Move completed work to archive, new work to active

• Refine skill files when outputs miss your standard

The system compounds. Every week you use it, the files get richer, the context stays more current, and the skill files get more precise. Three months in, it doesn't feel like using a tool. It feels like working with someone who actually knows your work.