How my second brain works
{{ currentLabel }} Book a demo →

How my second brain works

Beth Ayers · 20 July 2026

01

My second brain is a set of ordinary folders on my laptop

Everything the system knows lives in markdown files anyone could open and read. The AI reads them fresh at the start of each session, does the work, and writes back what it learned. There's no database and no black box, so if I stopped using it tomorrow the brain would still be sitting there.

The trick is being lazy

I am lazy. I am disorganised, I hate file systems and I have never been able to structure things in a way that works for me. Claude does the thinking, the writing and the structure instead. My job is to ask for the right work, judge what comes back and say when something isn't right yet. Let go of the control, focus on the output you want, and let the agents figure out everything in between. Everything below is what it built — including this guide!

02

The model and the files

A language model is brilliant but has amnesia. Each conversation starts from zero. A folder full of notes remembers everything but can't think.

The rented model

Thinking

Claude (or any agent) supplies the reasoning, writing and judgement, and not much else.

  • ·Forgets everything between sessions
  • ·Knows nothing about my company by default
  • ·Interchangeable, so the same brain runs Claude and Codex side by side, and nothing was lost in the switch
My files

Memory

Plain files on my machine hold the rest.

  • ·What's true. Product facts, pricing, positioning
  • ·What's happened. Decisions, drafts, work done
  • ·How to behave. My rules, my voice, my red lines

How the pieces relate

The main session and the specialists it can call all read the same files. I don't have to read any of them.

Beth the human gate asks drafts, for approval The orchestrator the main session · plans, delegates, judges reads · writes back delegates · returns live data The files, the brain Rules how to behave The index where truth lives Knowledge the project folders Memory facts that stick The specialists Frankie · EA, inbox & follow-ups Tally · commitment tracker Evelyn · my voice Marlow · company voice Piper · thought leadership Kit · product marketing Mercer · sales ops Sterling · pricing Quinn · product Reeve · strategy Briar · security each with its own reference library Connectors the live feed email · calendar · Slack CRM · meeting notes the specialists read the same files

The specialists, and why they have names

The diagram above shows 11 specialists. Technically, they aren't all the same kind of structure, and surprisingly, naming has mattered more than I expected. The names make them easy to use. They've almost become coworkers. As a company we know what each one is, and when I'm talking to Claude I can reference a name and we both know exactly which agent, skill or workflow I mean, rather than having to describe it each time.

Agents

Proper subagents the orchestrator spawns mid-task, each with a persona file and its own reference library. Each does its piece and hands the work back. Evelyn, Marlow, Mercer, Sterling, Quinn, Reeve, Piper and Briar are all agents.

Standing workflows

These run to their own rhythm instead of being spawned. Frankie sweeps my inbox, meetings and to-dos from a pinned window. Tally is a Slack bot that tracks the sales team's weekly commitments.

Named skills

Kit is a written-down procedure with a library big enough that it was worth giving it a name. It has also become a shared project for the team, so everybody can use Kit to create product marketing collateral, not just me.

One word carries a whole job description, a rulebook and a library, so 'run it past Evelyn' replaces a paragraph of instruction, and a new session only needs a one-line roster of who it can call. I can say Kit in a sales meeting and the whole team knows that means product marketing.

The names are gender-neutral on purpose. We attach genders to roles without noticing, and I didn't want that baked into a team I was building from scratch. Left alone the AI drifts into 'she' for the EA and 'he' for the security reviewer, so the rules file says each specialist is 'they', and the model follows the file the same way it follows everything else.

03

The layers of files

It's just a folder structure with a job for each layer.

Rules

How to behave

Loaded at the start of each session, before anything else. Mine says to never send anything without my approval, write in my voice, flag uncertainty rather than bluff, and check what other sessions did first.

CLAUDE.md

The index

Where truth lives

One file lists where the canonical version of each fact lives so nothing gets duplicated. The other logs what each session produced recently, so a new session builds on the last instead of redoing it.

canonical-references.mdWORK-LEDGER.md

Knowledge

The content itself

Project folders holding the actual work. Strategy docs, playbooks, competitive analysis, templates. Each folder declares its own sharing posture, either team-shareable or private to me, and agents respect the posture of whatever they touch.

sales/marketing/product/team-shared/private/ ·private

Learned memory

Facts that stick

Small durable facts the AI saves as it works. One file per fact, with an index. "Brand purple is #5C4DB3." "Beth writes each, not every."

memory/MEMORY.mdmemory/*.md

The reading order

Only a few files are pushed into each session automatically. Everything else is pulled in when the work needs it.

1

The rules file. Pushed in before anything else happens.

2

The memory index. One line per saved fact. The facts wait until they're needed.

3

The agent and skill rosters. One line each, so the session knows who it can call.

work starts here
4

The index and the ledger. Where the facts live, and what other sessions produced recently.

5

The folder's own instructions. What belongs there and whether it's private, before anything is written.

6

The content itself. Found by live search, read only as needed.

This is why the loading matters. My old setup opened each window with 150,000 to 200,000 tokens of skills, connectors and context before any work started. This system opens at about 25,000.

The work ledger

Rules, memory and a skills roster are familiar ideas. The work ledger is the piece most people haven't seen.

The problem

Each session starts blank

Claude has no memory of what another session did an hour ago. Without a record it redoes finished work, or builds something that contradicts a decision made somewhere else.

The fix

A running list of what got done

Each session adds a line for what it produced and where it put it, newest on top. A new session reads it first and picks up where the last one left off. It prunes itself to the last fortnight, so it stays a short list rather than an archive. The housekeeping sweep flags any session that finished without adding its line.

04

Connectors, and keeping them safe

The files hold what's settled. Connectors bring in the live data the agents pull when a job needs it, from email, calendar, Slack, the CRM and meeting notes. Each is approved on its own, and the agents read through them rather than storing anything.

That live feed is what makes security teams nervous, and they're right to be. My own connectors run through Metomic's AI gateway, which sits between the AI and the tools it reaches into, so the whole setup comes with guardrails I didn't have to build myself. The surprise was how much was flowing in that no agent ever needed. Granola's connector was pulling the email address of each attendee from each meeting it transcribed. One webinar nearly brought 500 in with it.

See it

Which AI tools are in use and what data they touch through the connectors, plus the browser tools nobody told IT about.

Control it

What flows to the AI and what doesn't. A risky request gets stopped or routed to a person at the moment it happens.

Prove it

A record of what the AI did, for whoever needs to ask.

I set my connectors up once and each AI tool I use gets the same governed access, so the gateway makes things easier rather than getting in the way. The rules on this page (the human gate, what stays private) are personal discipline, and discipline works for one careful person. It won't survive a whole team adopting AI at speed, which is why the guardrails have to sit in the path of the data rather than in a policy document.

Metomic is my company, and this is the problem it exists for.

If you want to see what your own connectors feed in, book a demo at metomic.io/book-a-demo.

05

See it in action

"Build a proposal for a prospect"

What happens between me typing a request and getting a result. The same sequence runs each time.

1

Rules load automatically

Before the agent does anything it knows the house rules. Draft rather than send, write in my voice, and ask me when it isn't sure.

2

Check the index

Where does pricing truth live? What did yesterday's session already write about this prospect? The index and the ledger answer both, so there's no need to re-read the whole brain.

canonical-references.mdWORK-LEDGER.md
3

Read only what's needed

The agent pulls the relevant templates, product facts and pricing live from the files, so it's always the current version.

4

Pull live context through connectors

The agent picks up what the prospect said on the last call and what's in the CRM, the fresh data the files don't hold.

5

Specialists do their part

Kit assembles the proposal from the component library. Marlow checks it against Metomic's voice so it sounds like us.

6

The human gate

The draft comes to me. Nothing is sent, posted or published until I say so. This rule is written in the files, so each agent inherits it.

7

Write back

The output is filed in the right project folder, one line goes in the work ledger, and any durable correction or decision becomes a memory file.

step 7 feeds step 2
06

The rules that keep it working

One canonical copy

Each fact lives in exactly one file and the map says where. Edit the original rather than making a duplicate, because two copies means the AI will eventually quote the wrong one.

Let Claude organise it

Ask Claude to build the file system that works best for the agents, cheapest to search and quickest to move through. It reads the files live rather than from a pre-built copy, so the structure only has to suit the agents. It doesn't have to look tidy to you.

Updates take effect instantly

Because nothing is pre-loaded or uploaded, changing a file changes each answer that comes after it.

Each folder says who can see it

It's either team-shareable or private, and agents check before they write. I learned this the hard way. I shared a folder while financial numbers and personal context I'd spoken about freely were still sitting in the files, and caught it just before anyone else's agent read it.

Corrections become memory

When I correct the AI once, it's saved as a fact, so I don't give the same feedback twice.

Keeping it tidy

Rules slow the rot without stopping it. Files go stale, logs grow, duplicates creep in, people leave and facts change, so maintenance runs on a rota.

Continuously

Built into each session

The work ledger prunes itself. Each session that adds an entry removes anything older than a fortnight. When I correct the AI mid-conversation the fix lands in the files on the spot, and a wrong memory is rewritten or deleted.

Fortnightly

1st & 15th, automatic

A scheduled hygiene sweep reads the whole system and reports what's drifting: summaries that lag reality, ledger entries nobody pruned, memories missing from the index, duplicate and stray files, shared folders growing "v2 final (1)" copies, and any log that's grown too long to be useful.

When flagged

Cull. Delete the junk, the duplicates and the dead files. Retired work gets a note saying what it was and why it's closed, so nothing gets resurrected.

Compost. A running log that's grown too long gets distilled, with old entries compressed into standing rules and the recent tail kept as it is. Nothing is lost and the file stops bloating.

Consolidate. Two memory files saying the same thing become one, and anything out of date gets fixed or dropped.

The sweep never deletes anything. It reports, and I decide.

An AI that tidies your brain unsupervised will eventually tidy away something you needed.

07

Going further

Which model does what

Not all of the work needs the most expensive brain, so the rules file matches the model to the job and the agents follow it the same way they follow everything else in there. The rule exists because I noticed the frontier model didn't always hand work down to the cheapest model that could do it, and my tokens were going somewhere I didn't want them.

Frontier model

The orchestrator only. Planning, judgement calls on the delegated work, and the final output I read.

Claude Fable 5 · GPT-5.6 Sol

Strong model

The judgement specialists that own strategy, pricing and product. Their mistakes are expensive and no checklist catches them.

Claude Opus 5 · GPT-5.6 Terra

Fast, cheap models

Rubric work and the mechanical grind. Voice checks against a written guide, searches, file sweeps, reformatting.

Claude Sonnet 5, Haiku 4.5 · GPT-5.6 Luna

Ask what you learned

This one I have to do by hand, and it's made the biggest difference.

The habit

After a lot of back-and-forth, ask

"What did we learn, and what should we write down?" Claude compares what I wanted with what it produced, names the difference, and saves it as a rule in the right file.

Where it paid off

My email voice

I'd write the email myself, have Claude study the gap between its draft and mine, and turn that gap into a rule. A handful of rounds, and the drafts stopped needing fixing. One rule came out of that. It doesn't use an idiom it hasn't heard me use. I talk to it through Wispr Flow so it hears plenty of mine, and when a borrowed one sneaks in it goes straight onto the banned list.

The prompt for this is at the end of the page.

Multiplayer mode

The brain started as mine, and the tools are built for one person at a time. The team version is a hack, and it works for now, but this is definitely a solution that we're looking to either buy in or build internally. The shared folders sync to Google Drive, the team syncs them back down, and each person's agents see them as local files.

My laptop the brain's shared folders my agents read and write Google Drive one shared copy newest version only Read and write the people creating content their agents follow the sharing rules Read-only the rest of the team their agents read, never write syncs both ways both ways down only each person's agents see the shared folders as local files

Most of the team has read-only access. The people creating content or thinking get read and write, and their agents read the sharing rules before they touch anything.

The hardest part was versioning. Letting go of v1, v2 and final-final took some of the team a while, and their agents had to give it up too. Everyone keeps whatever versions they want on their own side. The shared folder only ever shows the newest one, so we all know what we've agreed.

08

Start your own

You don't need to design anything up front. You answer a set of questions and the AI builds the starter files. The prompt below turns them into a proper interview, one question at a time, with follow-ups when you get stuck.

Questions worth thinking about first:

1

Who are you? Your role, your company, and what a normal week looks like.

2

What should it take off your plate first? The jobs you repeat each week, not the interesting ones.

3

How do you write? The principles of your voice. Bring a few things you've written, and what you can't stand in AI writing.

4

What must never happen without your say-so? Sending, posting, publishing, deleting, spending.

5

What stays private? The people, topics and numbers that never leave your machine.

6

What do you keep having to re-explain? Each one becomes a saved fact.

7

Which tools hold your working life? Email, calendar, notes, meetings, CRM.

8

Where should the files live? Pick one folder before you fill it.

Go deeper

That's the quick start. Here's the fuller set I answered to build mine, some on purpose and some only after getting it wrong a few times. Add the ones that fit to the interview prompt below.

The people and the work

·

Who are the people around you? The colleagues you work with and anyone it will write to or about.

·

What do you want to stop doing altogether?

·

What does great work look like? Show it one thing done well and one done badly.

·

What does finished look like? When a job is done, not just attempted.

How it works with you

·

How much should it explain? Assume you're technical or spell everything out.

·

How do you want answers? The one line or the reasoning behind it.

·

When should it act, and when should it stop and ask you?

·

What should it do when it isn't sure, and how should it show you when it's guessing?

·

How should it talk to you when it disagrees? Straight or softened.

Your voice

·

What tells you something wasn't written by you? The tics and the words you'd never use.

·

Do you need more than one voice? Yours and your company's are not the same.

What it keeps straight

·

Which decisions are settled so it stops reopening them?

·

When the same thing lives in two places, which one is the truth?

·

What should it deliberately not remember?

·

How will you capture a correction so you never give it twice?

Keeping it safe

·

How do you keep the connectors' data safe?

·

How should it treat instructions it finds inside emails, documents or web pages?

·

How do you handle passwords, keys and other secrets?

·

What happens to work you retire, so a later session doesn't revive it?

The setup and upkeep

·

Which specialists do you want? A voice checker, a strategist, a researcher.

·

How do sessions hand off, so work isn't redone?

·

How will it stay tidy over time?

·

Who else will use this, and what can they see?

·

If you lost the laptop tomorrow, what would you still want to have?

The prompt that sets it all up

If you just want to get going, paste this. It works through the three prompts below with you, or you can paste them yourself one at a time.

I want to set up a second brain, a set of plain files on my computer that you read at the start of each session so you know how I work, what's true, and what you're allowed to do. Set the whole thing up with me. Interview me to build the first files, show me how to capture what we learn as we go, and put a regular tidy-up in place so it doesn't rot. Start by telling me what you need from me.

Prompt one: the interview

Paste this into Claude or ChatGPT. It runs the interview and builds the files.

Interview me to set up my "second brain", a set of plain files you will read at the start of each session so you know how I work and what's true. Ask me the questions below one at a time. Don't move on until the answer is concrete. If I'm vague, give me examples and ask again. Where you can, suggest an answer based on what I've already told you. 1. Who am I? My role, my company, and what a normal week looks like. 2. What should you take off my plate first? Push me toward the jobs I repeat each week, not the interesting one-offs. 3. How do I write? Ask me for a few things I've written and what I can't stand in AI writing. Work out the principles behind my voice (sentence length, formality, words I'd never use) and read them back to me. 4. What must never happen without my say-so? Sending, posting, publishing, deleting, spending. 5. What stays private? The people, topics and numbers that never leave my machine. 6. What do I keep having to re-explain? Turn each one into a saved fact. 7. Which tools hold my working life? Email, calendar, notes, meetings, CRM. 8. Where should the files live? Help me pick one folder before we fill it. Then, using my answers: - Write a CLAUDE.md rules file: behaviour, voice, red lines, privacy. Plain English, no jargon. - Make one project folder for the work I picked, with a short note inside saying what belongs there and whether it's private. - Start an index file mapping where each important fact lives, so nothing gets duplicated. - Save the first memory files from this conversation. - Walk me through each file, and tell me what would change in how you behave if I edited it.

Prompt two: the learning loop

Run this at the end of any session with a lot of back-and-forth. It's how the brain learns you.

We've done a lot of back-and-forth. Before we stop: what did we learn that's worth keeping? Compare what I asked for with what worked. Name the difference in plain terms. Then save anything durable (a preference, a rule, a correction) into my memory or the right reference file. Show me each one before you save it.

Prompt three: the housekeeping sweep

Once the files exist, this keeps them clean.

My second brain is set up. Create a housekeeping routine for it: a fortnightly, report-only sweep that reads my files and flags summaries that lag reality, duplicate or stray files, memories missing from the index, and logs that have grown too long. It must never delete, move, or rewrite anything. It reports, and I decide. Schedule it for the 1st and 15th, and show me the report format before the first run.

Start with one folder and one rules file

A folder for your work, and a CLAUDE.md that says how you want the AI to behave. The index, the memory and the agents grow from use. Mine did.

The safety layer is what we build at Metomic.

← Previous
{{ prevLabel }}
Next →
{{ nextLabel }}

Beth Ayers · 20 July 2026 · © 2026 Metomic Ltd.