What is a second brain in this setup?

A second brain is a durable knowledge system outside chat history. Mine is two git repos with markdown notes, not a single AI tab. Personal holds life and home context. Engineering holds Skcript CTO work. Claude Code reads CLAUDE.md and noun folders each session, then writes decisions back so context compounds instead of resetting.

Why use two second-brain repos instead of one vault?

One vault mixed life and work until the wrong hat contaminated the next session. I split in July 2026 into Personal and Engineering with separate remotes and separate CLAUDE.md files. Home ops stay personal even when company services run on home hardware. Product strategy stays in Engineering. Correct hat beats a prettier graph.

What folder structure does the second brain use?

Each repo uses five parts: noun folders as the knowledge base, raw/ for immutable captures and conversation dumps, wiki/ for thin synthesis across folders, index.md as the map, and root CLAUDE.md for operating rules. Code projects may sit nearby with their own git history. There is no Inputs/Process/Outputs ceremony and no status-badge taxonomy.

What are noun folders in a second brain?

Noun folders are topic directories named for the thing itself—homenetwork, cloudflare, skcript-ceo, electronics—not abstract verbs like Areas or Projects. Detail lives in those notes. Future-you and the model both search for the thing, not a category invented at midnight. That is the real knowledge base.

What goes in raw/ versus wiki/?

raw/ is append-only ground truth: conversation snapshots and captures you never hand-edit. Compile from it; do not rewrite history. wiki/ is only for maps that span two or more noun folders after you have explained the same link twice. The wiki is a map, not a second copy of every note.

How does Claude Code use CLAUDE.md in a second brain?

Root CLAUDE.md defines the hat, filing rules, and commit habit for that repo. Folder-level CLAUDE.md files add local scope when useful. Each session starts by reading those rules and the relevant noun notes. After work, the model files what was learned, updates links and index.md if needed, then commits. Chat is disposable; the note is not.

Does a second brain replace a human assistant?

No. This system replaces context rebuild between AI sessions. It does not own relationships, make phone calls, or care about a deadline you never wrote down. If an insight only lives in a closed chat tab, you do not have a second brain—you have a search box with amnesia.

How should I start a second brain like this?

Split by hat first if work and life both need notes. Name folders after nouns. Keep raw immutable and compile carefully. Make the model write the note after every real session. Keep wiki thin. Skip quarantine folders, morning vault walks, and graph cosplay until a real failure mode appears. Structure exists to lower filing friction, not to perform knowledge management.

How I Manage My Second Brain (Two Repos, Claude Code)

I manage my second brain as two git repos and a rule that notes outlive chat. Personal is the life hat. Engineering is the Skcript CTO hat. Every few months Twitter rediscovers a three-folder, one-evening system that will replace your assistant by December. I run a second brain every day. It looks almost nothing like those posts.

If you already use Claude Code as an always-on copilot, this is the knowledge layer that stops you re-briefing it every morning.

Why did I split my second brain into two repos?

I used to keep everything in one vault. Life, home lab, FeatureOS strategy, server inventory — one tree, one context window, one mess of mixed hats.

In July 2026 I split it.

Personal is the life hat: home network, Home Assistant, reMarkable, music, devices, weekend projects, this site. Engineering is the Skcript CTO hat: Cloudflare, Coolify, server management, SupportWire GTM notes, EA handoff playbooks, product maps.

They sit next to each other on disk:

~/Developer/Personal
~/Developer/Engineering

Separate git remotes. Separate CLAUDE.md. Separate rules about what may be filed where.

That split is the whole product. Home iron can run a company service (Rybbit on the Mac mini) without product notes living next to AirPods research. The work angle goes in Engineering. The house ops angle stays in Personal. Cross-links when needed. No mega-graph of everything I have ever thought.

I also do not fall for Twitter hype when a thread claims three folders replace a team. The durable part is hat separation and write-back, not the demo.

What folder structure does each second-brain repo use?

Each repo has the same boring shape on purpose:

Noun folders — the actual knowledge base. Named for the thing: homenetwork/, cloudflare/, skcript-ceo/, electronics/. Not verbs. Not abstract buckets like “Projects” or “Areas” unless the thing itself is a project.
raw/ — immutable sources. Conversation snapshots from Claude sessions. Captures and exports. Append-only. Compile from it. Never rewrite history because last week’s self was wrong.
wiki/ — thin synthesis only. Maps that span two or more noun folders when I have already explained the same connection twice. The wiki is a map, not a mirror. Detail stays in the noun notes.
index.md — master map. Entry point when a human or a model is lost.
Root CLAUDE.md — operating rules for that hat: what this repo is, what it is not, how notes get written, when to commit.

Some folders are also code (site/, small services). Those keep their own git history when they need it. The knowledge repo links to them; it does not pretend to own them.

That’s it. No Inputs/Process/Outputs/Feedback. No status badges. No “CANON / STALE / FILL” legend. Clear title, what it is, what to do. Limit your options and do them well applies to note systems too.

How does the model own the notes?

The important rule is not the folder names. It is who writes after a session.

When something is learned — a firewall decision, a loan account number, a GTM call, a contradiction with an older note — the model writes it into the right noun folder, links with relative markdown, and updates index.md if the map moved. Then commit. Prefer push. Don’t force-push.

Chat is disposable. The note is not.

That flips the usual AI failure mode. Without a vault, every session starts with you re-explaining yourself. With this setup, the session is a compile step: read the hat’s rules, read the relevant noun folder, do the work, write back what changed.

I also keep a small self folder under Personal — how I work, how I want agents to write for me. Not a personality cosplay document. Just enough so the model stops reinventing my voice and my taste for simple pages.

What is raw/ for in a second brain?

raw/ is ground truth that never gets cleaned up by hand.

Session history lands under raw/conversations/.... Clippings and exports under raw/captures/. If I need the original dump six months later, it is still there. If the compiled note was wrong, I can re-compile. I do not “fix” the capture.

Engineering’s raw is mostly conversations today. Personal has more captures. That is fine. Structure exists so friction stays low when something should be kept, not so every thought becomes a ceremony.

Why keep the wiki deliberately small?

People hear “wiki” and build a second copy of every page the model already wrote under a noun.

Mine only triggers when a theme spans folders and I have paid the explanation tax twice. Home infrastructure across network + Mac mini + Home Assistant. Product map across services. Reading pipeline across reMarkable + Readeck + Kindle.

One page. Links out. No duplication contest.

What I refuse to do in my second brain

I do not run a morning agent that walks the whole vault and emails three lines of “what changed overnight.” Sounds nice. I have not needed it yet. Committing after real work beats scheduled maintenance theater.

I do not treat the second brain as a replacement for a human assistant. It replaces context rebuild. It does not own a relationship, pick up a phone, or care about a deadline you never wrote down.

I do not merge Personal and Engineering to get a prettier graph. The graph is not the goal. Correct hat, low filing friction, notes that survive a closed tab — those are the goals.

I do not reorganize because a thread said three folders is enough. I reorganized once when the unified vault started lying to me about which hat I was wearing. That was enough.

How does a normal second-brain session work?

Open the right repo. Life problem → Personal. Company problem → Engineering.

The root CLAUDE.md loads. Topic folder rules load when relevant. I point at a note or drop a capture. We work. Something is decided. The model files the decision next to the thing it is about, not into a generic “Outputs” bucket. Index updated if needed. Commit.

Tomorrow I do not re-brief who I am or what the home network looks like. The folder already knows.

How should you steal this second-brain system?

If you want a second brain that actually compounds:

Split by hat before you split by folder fashion. Work and life contaminate each other fast.
Name folders after nouns. Your future self searches for the thing, not the abstract category you invented at 11pm.
Keep raw immutable. Capture freely. Compile carefully.
Make the model write the note. If it only lives in chat, you do not have a second brain. You have a smarter search box with amnesia.
Keep wiki thin. Synthesis maps, not a second copy of reality.
Skip the ceremony until pain shows up. Quarantine folders, daily briefings, graph cosplay — add them when a real failure mode appears.

I built this for me and for the agents that work with me every day. Not for a demo. Not for a screenshot of a knowledge graph.

Two folders. Two hats. Notes that compound because someone — usually the model — actually files them.

How I manage my second brain with two git repos