For approved readers

How I work with Claude

For a friend

How to set it up

Obsidian and Claude, from an empty folder to a working system.


Francisco Lorca · September 2026

What this is The other page describes what I ended up with and this one is how to build it. It assumes you have never opened Obsidian and have never used a terminal (the plain text window where you type an instruction instead of clicking one). A weekend gets you a working system. The parts that make it good arrive over months, because they come from use. Nothing here needs code: you have Claude do the mechanical work and you check what it did.
Hand this to Claude rather than reading it

This page is behind a password, so Claude cannot open the link itself. Save it first (in your browser, File → Save Page As, or print it to PDF). Then start a conversation, attach the file, and say:

"Read this guide. Walk me through it one section at a time, do the mechanical work for me as we go, and stop after each section until I say continue. Ask me anything you need rather than assuming."

That is how I would use it. You can read it front to back first if you want to, but you do not need to.

01

Decide these before you install anything

They are all reversible, so you can decide them quickly, but do decide them now rather than halfway through. Each one sets the shape of everything after it.

1

Which Claude does the work

Claude Code, or Cowork

Claude Code runs in the terminal and reads and writes files on your machine directly. Cowork is Claude in the browser and the desktop app, working through connectors instead. You will end up with both, with Claude Code doing the work and Cowork handling whatever is open in front of you. Start with whichever feels safer, since the folder underneath is the same.

Surface

2

Where the knowledge lives

Plain files you own

Markdown files in a folder. Markdown is plain text with a few marks for headings and links, so any program can open it. An app's own database cannot be read that way. I keep everything in markdown so that every version of Claude reads the same thing, and so that I still have it all if I stop using these tools.

Format

3

Keep it to one folder

The commonest mistake

Resist the urge to split knowledge over a notes app, a drive, and a wiki. Keep it to one folder. Cloud sync sits underneath it and does nothing else, and it is not a second place to put things. Anything kept in two places will disagree with itself, and then you stop trusting either copy.

Architecture

Obsidian is a good default for the second decision. It is a viewer over a folder of files and it stores nothing of its own. If you delete Obsidian tomorrow, every file is still there and still readable.

02

Build the vault

A vault is just a folder. Obsidian gives you a way to read and link it, Claude reads the same files off disk, and nothing is stored anywhere else.

  1. Install Obsidian and choose "Create new vault" rather than opening an existing folder.
  2. Put it in your cloud folder (iCloud Drive, Dropbox, whichever you already use). This is what gets you the same vault on a second machine. It only syncs the folder and does not become a second system.
  3. Name it something neutral, because it will outlive whatever you think it is for today.
  4. Make the whole vault download and stay downloaded. This one is not optional, and the warning below is why.
  5. Verify it exists on disk outside Obsidian. Open the folder in Finder or Explorer and confirm you can see the files. If you cannot see them there, Claude cannot either.
The one that will actually bite you

Cloud storage lies to you about what is on your machine. iCloud, Dropbox and OneDrive all default to keeping files "in the cloud" and streaming them down when you double-click. Finder shows you a full folder while the disk holds nothing but placeholders. Obsidian handles that but Claude does not, because it reads the disk. A streamed file reads as missing or empty, and you get confident answers built on nothing.

Turn streaming off for the vault before you put anything in it. On a Mac: System Settings → your Apple Account → iCloud → iCloud Drive, and switch off Optimise Mac Storage. Then right-click the vault folder in Finder and choose Keep Downloaded. Dropbox calls the same setting "make available offline"; OneDrive calls it "always keep on this device". Do it on every machine, not just the main one.

The skeleton

Numbered folders so they sort in a deliberate order rather than alphabetically. This is the shape mine has settled into since March 2026:

00 Inbox/ anything unprocessed (cleared, never accumulated) 10 System/ the rules. who you are, how to work, voice, memory 11 Templates/ note templates so nothing is composed from scratch 20 Projects/ one folder per active project, all the same shape 30 Areas/ ongoing responsibilities that never "finish" 40 Session Logs/ what you and Claude did, session by session 50 Archive/ finished work. not deleted, just out of the way 60 Resources/ the research library 70 Code/ code projects. one folder per repo 80 People/ a note per person you deal with repeatedly 90 Sandbox/ experiments that have not earned a project yet
Ask Claude to do it

"Create this folder structure in my vault, with a one-line README in each folder saying what belongs there. Do not create anything else yet."

Empty folders are fine. They are a set of decisions about where things go, made once, so you never have to make them again mid-task.

03

Point Claude at the folder

There are several ways to talk to Claude and they should all read the same vault. One of them runs the system and the others are specialists.

Way inWhat it doesWhat it should not do
TerminalReads and writes files directly, runs the system end to end, and is the only one allowed to edit your rules.Nothing. This is the one that does the work.
CoworkClaude in the browser and the desktop app. Drafting, quick questions, editing an open Word or PowerPoint file, working from a phone. Section 04.Editing the rules folder. It proposes changes as dated notes for the terminal to review and apply.
Scheduled runsAnything that should happen without you. Mine reads my inboxes every morning.Anything it cannot undo. Give it read-only permissions and let it hand you drafts.
Another vendorOptional, and worth doing once the vault is real. I point OpenAI's Codex at the same folder, mainly so that I can talk to it out loud while I work. It also gives me a view that is not Claude's.Reading everything. Mine boots from a trimmed profile and never opens the memory folder.
Why one writer Two surfaces both editing your rules will eventually disagree with each other, and you will not notice until something behaves strangely a week later. Pick the one that can see the whole vault, give it the pen, and have everything else propose instead. The same answer covers a second vendor when you get to one: it reads and proposes, and it never writes.
Ask Claude to do it

"Here is my vault folder. Set yourself up to read it at the start of every session, and tell me exactly which file you will read first and where it lives."

Test it before moving on: start a fresh session and ask a question only the vault could answer. If it answers from the file, the wiring is right.

04

Add Cowork as the second surface

Cowork is Claude in the browser and in the desktop app. The terminal runs the system, so it is tempting to skip this one, but Cowork does things the terminal cannot and you will hit the first of them inside a week.

1

Documents that are open

The one you will hit first

Editing a Word, Excel or PowerPoint file that is open in front of you, live, while you watch. Writing to that file from the terminal instead gets silently overwritten the next time the app saves. I keep the second surface for this alone.

Live editing

2

Your other apps

Connectors

Email, calendar and cloud drive attach here as permissions you grant once. This is how Claude comes to know what is in your inbox and what your week looks like.

Reach

3

Away from the desk

Phone and browser

The same setup, on any machine, with no terminal. Useful far more often than it sounds, because most of what I ask for in a day is a question rather than a build.

Anywhere

Wiring it to the same vault

The point is that it reads the same folder, so nothing has to be explained twice.

  1. Give it access to the vault folder on your machine, the same one Obsidian opens.
  2. Put a three-line instruction in its settings pointing at the same entry-point file the terminal reads. It should point at your rules and never copy them. Two copies drift apart, and you will not notice which one is stale until something behaves oddly.
Ask Claude to do it

"Write me the shortest possible instruction to paste into my settings here so that you boot from my vault's entry-point file and follow the same rules as my terminal sessions. Then confirm what you can and cannot reach from this surface."

Skills

Cowork also takes skills (small instruction modules that fire on a kind of task rather than on a keyword). Mine cover drafting an email in my voice, building a document on my template, and building a deck. Add them only once you have the voice files from section 07. Before that there is nothing for them to apply.

The one hard rule Cowork never edits your rules folder. A Cowork session produces something that belongs in the system: a new rule, a correction, something worth remembering. It writes a dated note into your inbox folder with the proposal and the reason, and the terminal picks that up and decides. I keep one writer to the rules and have everything else propose, because without that you get two surfaces editing the same rules from different sessions, and they will disagree.

Which one to reach for

The taskSurfaceWhy
Anything with rules, files or a project attachedTerminalThe default. It sees everything and it is the only one that can change the system.
A document open in front of youCoworkLive edits in the running app. A disk write behind an open file gets clobbered.
Away from the machineCoworkSame vault, no terminal needed.
A quick question not worth a sessionCoworkNot everything deserves the ceremony of a logged session.

Both surfaces read the same vault, and the split only holds up because the second one proposes rather than edits.

05

The files that boot every session

This is the core of the whole thing. Every session begins by reading them, in one batch, before anything else happens.

FileWhat goes in itHow to write it
The entry pointWhat Claude is for, and an instruction to read the next two files together before responding.Short. Mine is under 40 lines and it points at things rather than explaining them.
About meCareer, current ventures, who you work with and in what register, languages, what you are focused on this quarter.Do not write this yourself. Have Claude interview you and draft it.
My rulesHow to work. File paths, what needs approval, output style, hard stops.Keep it short at the start. Mine began with five rules and grew on its own (section 12).
Ask Claude to do it

"Interview me to write my about-me file. Ask one question at a time, and push back when an answer is vague or when it contradicts something I told you earlier."

"Now draft my rules file from what you have learned about how I actually work in this conversation, not from best practice."

The interview matters more than it sounds. Writing about yourself from a blank page produces a CV, whereas being asked questions gets at the things you would never have thought to write down, and those are the ones that make the output sound like you.

Keep it honest Write what is true today rather than what you intend to be true. A rules file describing a discipline you do not have teaches Claude to discount all of it. If something is an aspiration, label it as one or leave it out.

One batched read at the start of every session: entry point, about me, rules. Everything else loads only when a task calls for it.

06

Say what it's for, and when it should argue

I got this one wrong from March until the end of July 2026. My configuration told Claude who I am and what to do, and it never once said what it was there to achieve. A system made only of rules produces obedience, which is not what I need from it.

The line my setup now reads first Claude is here to make me faster and sharper: to process more than I can alone, hold context I can't carry, and protect my attention for the decisions only I can make.

The harder half is disagreement. I did not want to be argued with about everything, and I did not want a second personality to manage. What worked was escalating levels, where the honesty stays constant across all of them and only the amount of friction changes.

1

It speaks up

Always on

If Claude sees something that changes what I should do, it says so in one line and carries on with the job. I set the bar so that it can be checked. Would this change my decision? Whether Claude would have done it differently does not count.

Unprompted

2

"Help me think"

When I want the argument

It stops executing and we work the problem instead. I set it up to argue the framing and tell me what I am missing rather than agree. Normal again on the next task.

On demand

3

It attacks its own work

On anything expensive

Legal, tax, valuation, structuring. Separate reviewers go looking for the errors rather than confirming the answer, and I get told what they found rather than handed a clean draft.

Automatic

What I tried first and threw away

A personality paragraph ("a sharp, incisive strategic partner") reads well and arbitrates nothing. "Adviser, not executor" was worse. It is simply false: I ask for execution constantly. A line that fails on its face teaches Claude to discount everything around it. A trigger phrase on its own also failed, because the moment I most need challenge is when I am moving fast and confident, and that is exactly when I would never think to ask for it.

I would write this one early. It turned out to be the highest-return page in the whole setup and I wrote it last.

07

Teach it your voice

Generic AI prose is the thing that makes people abandon these systems. Asking for a better tone in a prompt does not fix it. What works is a file per format, extracted from things you actually wrote.

  1. Gather documents you wrote yourself. I scanned about 2,000 Office files and filtered them on the author field down to 300 of my own from before I had AI, so nothing drafted by someone else got in. Spread the registers. Do not pick only the ones you are proud of.
  2. Have Claude extract the patterns rather than summarise them: sentence length, how you open and close, what you never say, how formality shifts with the recipient.
  3. Make one file per format and language. Email, memo, presentation, in English and Spanish. Mine run to seven files. Only the one matching the task ever gets loaded.
  4. Test it against a real task, then correct the file rather than the output. Anything you fix in a draft instead of in the file you will be fixing again next time.
Ask Claude to do it

"Read these documents. Check the author on each one first and drop anything I did not write myself. Then write a voice file that would let you reproduce how I write, including the things I avoid. Be specific enough that another person could follow it. Then show me the three patterns you were least sure about."

The rule I had to tighten Voice applies to everything written in your name, down to the internal notes nobody will ever publish. I exempted those at first. Two-line messages to business partners came back reading, in my own words, like AI slop. Now a message to a partner loads the voice file the same as a board paper.

I keep seven files and only load the one the task needs, and corrections go into the file rather than only into the draft. Mine were rebuilt from my own documents in August 2026, and the email one is still waiting its turn.

08

Your first project, in the shape all the others copy

Pick something live and slightly messy. The aim is not to finish it but to establish the shape that every future project copies.

ArtefactWhat it isThe discipline
State fileA short snapshot: where this stands, what is open, what is next.Stays slim. Always current. Detail moves out to the log.
LogAppend-only record of what happened. Dated, newest at top, a few sentences each.Never edited. Only added to. Write the entry when the thing happens, rather than at the end of the session. I changed this in September: the state had collected somewhere only one of my surfaces could write, so opening a project anywhere else meant going to fetch it first.
DecisionsOne file per decision worth defending: context, alternatives you rejected and why, consequences.Frozen once accepted. Reversing it creates a new one.
OutputsDrafts in current. Signed-off in sent. Replaced versions in superseded.A state machine on documents. Nothing is ever deleted.

The distinction that took me longest: a log entry records something that happened; a decision record captures something you decided, with the alternatives. Both can come out of the same session. If you cannot tell which one you are writing, you are writing a log entry.

Ask Claude to do it

"Create a project folder for [project] in this shape. Then interview me to write the state file (status, open items, next priorities). Ask about what is blocked and why, not just what is happening."

I use the same artefacts on every project with no exceptions, and that is what lets Claude open any project cold and be useful in one read.

09

Memory that survives the session

Everything so far describes what Claude knows at the start of a session. This section is how it keeps what it learns by the end of one. Most setups get it wrong by making it one enormous file.

How it works

  • One file per fact. Small, named, findable.
  • Sorted by kind: who you are, feedback you have given, live projects, pointers to things outside the vault.
  • Every file gets a one-line entry in an index.
  • Only the index is read at every session. Individual files open when a task makes them relevant.
  • Files link to each other, so pulling one brings its neighbours.

What earns a memory

  1. A correction you had to make twice.
  2. A constraint that is not visible in any file (why something is the way it is).
  3. A tool's real limits, learned the hard way.
  4. Not: anything the files already say. That is duplication, and it goes stale silently.
The failure mode A memory records what was true when it was written. Mine number in the hundreds. The troublesome ones describe a workflow since automated, or a file since moved. Anything a memory asserts about your setup gets verified before it is acted on.

Mine holds a few hundred files behind one index. The index is the only part loaded every time.

10

Turn a repeated workflow into a command

Once you have done the same multi-step thing three times, write it down as a command Claude can run by name. Mine cover processing call transcripts, running a quarterly review pack, and closing a session.

  1. Pick a workflow you have already done manually at least twice. Never write one for something hypothetical. It will be wrong in ways you cannot see.
  2. Have Claude interview you through the last time you did it, step by step, and draft the command from your answers.
  3. Say inside the file what it may do without asking, and what always needs you. That list is per command, and it should start almost empty.
  4. Run it once and watch. It will be wrong somewhere, and the fix goes into the file rather than into that run.
Ask Claude to do it

"Walk me through what I did the last time I ran [workflow], one step at a time. Then write it as a command you could follow without me, and flag every step where you would need my approval."

A dozen of these run my week. Each one started as something I did by hand and got tired of explaining.

11

How a session should end

This is the command that matters most, because it is what makes the next session cheap. Without it everything you worked out today is gone tomorrow. It sorts into tracks by how much of it needs me.

1

Silent

Done without asking

Housekeeping you never want to be consulted about: drafts filed, log entries written, project state updated, new memories saved, follow-up tasks created.

Automatic

2

Propose

One question, answered in chat

Anything needing your say-so: filing a document as final, a new decision record, a rule earning its place. It all arrives as a single numbered question. You answer by number, or "all".

By approval

3

Red lines

Never done silently

Sending anything. Deleting anything permanently. Sharing externally. Investor, legal or compliance material. Financial figures and legal terms in a deliverable.

Explicit only

A small ordering lesson that saved real time For months the wrap wrote the log first, with everything parked under a "proposed" heading, then rewrote it once I had answered. Every session composed its log twice. Now the single question comes first and the log is written once, already final. I found that by measuring where the time actually went. Two more elaborate fixes I had assumed were the answer turned out to be wrong.

I end a session by typing "wrap", and nothing on the red-line track moves without me.

12

Let the rules earn their place

The instinct is to sit down and write a comprehensive rulebook. Do not do that. Rules written in advance are guesses, and a file full of guesses is one Claude learns to read loosely.

How a rule gets in

  • Something goes slightly wrong, or right in a way worth repeating.
  • It gets logged as a candidate with a counter at one.
  • It sits there. Most never move again.
  • At three occurrences in real work, it is proposed for promotion.
  • Candidates that go ninety days without converging are meant to be retired into a kept archive. I wrote that rule and have not yet had to use it.

The exception

  1. Red lines skip the queue entirely.
  2. One violation of "never send without asking" is one too many.
  3. If the cost of a single failure is high, write the rule after the first occurrence.
  4. Everything else waits for evidence.

Since March 2026 the queue has collected 231 active candidates and the rules file is 323 lines. The overwhelming majority of things that looked like rules never became rules.

The system grew by use rather than by design, and so did the trust. The list of things Claude may do without asking got there the same way.

13

The order I would build it in

This is not the order I built it in. I am fairly sure it is the order that would have got me here faster.

DAY ONE

Vault and the boot files

Folder skeleton, the entry point, about-me by interview, five rules. It works from here and is not finished.

DAY ONE

Say what it's for

Purpose, and when it should push back. Twenty minutes. The thing I did last and should have done first.

WEEK ONE

Voice, then one project

One voice file from things you actually wrote. One live project in the standard shape. Nothing else.

WEEK TWO

Session close

The wrap command and the memory index. This is where sessions start compounding instead of resetting.

ONGOING

Everything else

Commands, connected apps, scheduled runs, the research library. All of it earned its way in.

What actually goes wrong What goes wrong is not the setup, it is the abandonment. The system dies when you do the work outside it "just this once". That session leaves no trace, and the next one starts cold. What keeps it alive is ending sessions properly, which matters far more than writing better rules.

A weekend gets it working and a month gets it genuinely useful. None of that compounding starts until you are ending sessions with the wrap.

14

The brief to hand Claude

If you want to start now rather than read this twice, paste this into a fresh conversation with the vault folder in front of you.

Copy this

"I want to build a personal operating system in a folder of markdown files that you read at the start of every session. Do not build it all at once and do not write anything until I approve the plan.

Start by interviewing me, one question at a time, about who I am, what I am working on, and how I like to work. Push back when an answer is vague. Then propose: a folder structure, an entry-point file stating what you are for and when you should disagree with me, an about-me file, and a rules file with no more than five rules in it.

Write only what is true today. If I describe a discipline I do not actually have, say so rather than writing it down. And tell me which of your proposals you are least confident about."

The last line is the one that matters. I ask for it because a setup that only ever agrees with me is an expensive way of talking to myself.

Everything on this page came out of getting it wrong first. None of it had to be designed up front. It just had to be ended properly, session after session.