Build your second brain in 7 steps

Most second brains are docs graveyards. Beautifully organised folders nobody opens, clever chat histories nobody files, notes that never become work. This is the playbook I actually run, built across two companies, mistakes included.

Each step has one job. Do not move to the next until the current one is done. The discipline of closing one loop before opening the next is what produces the system. The content is downstream.

The 7 steps

1 · Setup
open the workspace

2 · Direction
one clear objective

3 · Externalise
dump everything

4 · Vault
memory outside projects

5 · Sources
plug in connectors

6 · Cadence
brief it like a teammate

7 · Output
decisions, then automate

1 · Setup. Open Claude Cowork. One workspace, one project. The AI lives next to your files, your calendar, your tools, not in a separate browser tab. This is the substrate everything else sits on. Do this: create a Cowork project and point it at the folder where your work lives.

2 · Direction. Set a clear objective. One sentence: what is this brain for? “Run my company end-to-end as a solo founder.” “Be the second brain for our growth team.” The objective drives the project instructions and frames every brain dump that follows. Do this: write the objective, then write the project instructions from it.

3 · Externalise. Brain dump everything in your head. Frameworks, decisions, learnings, in-flight contexts, current people, current playbooks. Time-ordered, raw, ugly. This is the baseline the brain validates against forever. Skip it and every later layer drifts. Do this: one long voice note or doc. Do not edit. Do not structure. Just dump.

4 · Vault. Connect Obsidian, outside the project folder. Your vault is operator-scoped, not project-scoped. It travels with you across projects, employers, and ventures. Let the sync skills handle frontmatter, tags, links, and inbox processing. Do this: create the vault outside any project folder and install the sync skills.

5 · Sources. Plug in your connectors. Email, calendar, Slack, project tracker, analytics, CRM, drive. Each connector turns one more part of your operating reality into something the brain can read. Do this: connect the two or three sources you already check daily. Test that the brain can read them. Do not wire everything on day one.

6 · Cadence. Treat it like a team member. The brain stays current the same way a teammate does: by being kept in the loop. Forward important emails. Drop meeting minutes after every sync. The second brain dies from missing rituals, not missing notes. Do this: book a recurring weekly slot and brief the brain like you would brief a chief of staff.

7 · Output. Produce use cases, then automate. Now the brain is ready. Use it to turn loops into decisions: analysis, specs, documents, models. Once a loop repeats, schedule it. The mechanics run on a timer while you keep the judgement. Do this: run your first three loops manually, then schedule the ones that repeat.

When all seven close, you have a second brain that ships ranked decisions on a cadence. Not a docs graveyard. Not a clever chat history. A brain that absorbs context like a teammate, produces filed work, and runs on calendar discipline rather than willpower. Anything less is a hobby with a markdown extension.

How to think about each pass

The iteration mindset is not “get it right”. It is “close one loop, learn, then open the next”. The operator who finishes a stage in a week ships a working v1. The operator who polishes everything in parallel for a quarter ships nothing.

One version answers one question. If a stage produces three findings, you have quietly run three stages, and none of them closed cleanly.

Fresh session per stage. Open a clean AI session for each new step and paste only the prior stage’s closing artifact as input. Context drift inside one long chat is what kills the loops.

Time-box at one week. If a stage is not closing in a week, the scope is wrong. Reduce, do not extend. The week-long box forces simpler v1s, which is a feature, not a bug.

The next version sharpens, it does not restart. v2 is not “redo v1 better”. v2 is “v1’s finding is the starting point, what do we now ask?” If you find yourself reopening v1 inside v2, the pipeline is broken.

Close ugly beats polish open. An ugly closed v1 always beats a beautiful open v3. If a closing artifact exists, you have leverage. If three stages are 80% done, you have nothing.

The deliverable is the filed decision. A clever chat reply that does not become a scored backlog item is wasted work. The brain ships filed decisions, not insights. That is the only output unit that compounds.

What I did wrong along the way

Real mistakes, made building this system across two companies. Each cost weeks. Each is now a guard-rail in the seven steps above.

I scaffolded the vault inside the project folder

Felt natural. Wrong on day one. A few weeks in I could not share the project folder with a teammate without leaking my private thinking. The fix: the vault moved outside every project. Project folders stayed operational and shareable. The vault travels with me.

I built the vault before the action layer

I polished folders and frontmatter, and three months in noticed none of my findings had become filed tickets. A beautiful docs graveyard. The fix: every finding now compresses to a one-line problem statement, scored, filed, owned. No exceptions.

I wrote standards for what the tool already automates

My first "team standard" documented formatting the skill already handled; the human-facing part was a footnote. The fix: every standard answers one question, what changes about human behaviour because this exists? If nothing, the doc gets deleted.

I answered new questions ad hoc in chat

Six discovery passes produced zero filed tickets. The thinking was fine. The pipeline did not exist. The fix: a new source is a discovery loop, not a one-shot reply, and the loop ends with filed problem statements.

I had no promotion ritual on the calendar

"I will promote when I have time" meant the inbox was permanent. The fix: a recurring Sunday slot, about thirty minutes, with a hard close: inbox empty, keepers graduated, junk pruned.

I installed skills inside the project folder

Same mistake as the vault. A skill is an operator capability, not a project asset. The fix: skills live at the operator-scoped path, never inside a project tree.

I kept one vault for everything

Work, personal, side projects, all wiki-linked. Elegant until a personal entry bled into a work search. The fix: one vault per active venture plus one durable personal vault. Cross-vault links do not work, and that is a feature.

I filed engineering tasks before discovery had closed

A [LATER] tag is still in the queue, and it still costs working memory. The fix: premature follow-ups stay as notes in the parent task. A ticket is filed only when the discovery that justifies it has closed.

The test

The brain is the system that ships ranked decisions on a cadence. Build it top-down. Close one loop before opening the next. Let the team stand it up without you. That is the only proof it is real.

This is the same system The Brain installs inside a company: one memory, wired to the tools the team already opens, run on rituals instead of willpower.