This one is a little morbid, and I want to be honest about it from the start.
I think a lot about the day I am going to die. Not in a dark way, so please don’t worry about me as you read this. I find the topic fascinating, the way most people do once they let themselves look at it. What happens on that day? What happens after? Is it all black, all white, a new world?
My own curiosity is less about the afterlife and more about this world. I look at it from two angles.
The first is mercy
I have lived with a lot of pain: chronic fatigue, and IBS bad enough that it makes me scream when it flares. I am recovering, and I work hard to keep it out of most of the hours in my day. But I would be lying if I said the thought of rest never felt merciful.
The second angle matters more, and it is preparation.
There is the preparation for the day itself. What the funeral looks like, how people are welcomed, which notes I leave behind for each person, what goes to my wife and what goes to my son. I want all of that in writing, so nobody is left guessing. Honestly, that part is easy.
The harder part is what it does to me when I think it through. It scares me back to reality.
Not just because of my wife and son, but because of everyone else it would touch: the instructors helping me work through the pain, the regulars at the coffee shops and restaurants where I have made friends with almost everyone, the families at my son’s micro school.
Thinking about them makes me want to be a better version of myself, even on the days the pain is loud.
And being better has a practical half.
While I am here, how do I improve the lives of the people around me, financially and otherwise, in a way that keeps going when I am not?
That is where the computer comes in.
Here is the question I ask myself. If I went away tomorrow, could someone sit at my computer and run things the way I do? The way I think things through, the tools I reach for, the problems I solve every single day?
If the answer is no, then whatever I have built on that computer dies with me. A writing business, a digital product business, a paid newsletter, consulting. It doesn’t matter what you call it. It is gone.
This is not a morbid invention of mine. It is the test buyers already use to value a company beyond the numbers on the spreadsheet. If the people who operate the business were to die, would the business survive? If it wouldn’t, it has no real value.
It is also why consulting businesses are among the worst businesses you can own. They cannot be fully automated, so they cannot be valued properly. A real business survives the death of its operator.
I know this sounds far-fetched. Why think that far ahead?
Because it can happen.
And if anyone relies on your writing income, whether that is your spouse, your kids, a friend or a business partner, the day you die is not the hard day. The hard day is a week later, when they wake up to your financial absence as well as your emotional and physical one.
Let me put this another way, because it sounds less grim.
Your death is also a holiday.
If the business needs you alive all the time, it needs you active all the time. There is no room for you to be inactive, to not show up, to take a week off. You have to walk into the factory every day and pull the levers with your own hands, or nothing moves.
A business that cannot survive your death cannot survive your holiday either. Build it for the first and you get the second for free.
Why the terminal, Claude Code and a Second Bain Operating System are the answer
So how does a writer, someone who is not technical, build that kind of business?
The terminal. Let me explain why, because it is not the obvious answer.
The browser was built for humans to share documents
The browser and the internet came about so that people in one office could share documents with people in another, anywhere in the world, and have them look the same.
That consistency is what the browser is good at.
Then the documents became dynamic.
They started pulling data from databases: tables of contacts, employees, orders. Documents turned into applications, with a front end you see in the browser and a back end that feeds it. Product owners and designers arrived, because now it mattered not just what loaded but how people clicked through it.
All of that was designed for one user. A human.
How we landed at Claude Code and why people are leaving the browser chat window for the terminal
At first there was no app at all.
GPT-3 proved that making a model bigger made it smarter, but you could only reach it through an API, a back door for developers. You sent it text, it sent text back. Request and response. It’s like the early web: the plumbing existed, but only a few people could use it.
Then, in late 2022, ChatGPT put that same request and response inside a chat window, and it proved that ordinary people wanted intelligence in a browser based application. Claude followed. And because the AI companies assumed humans would be the main users, they built what humans had always been given. An app. Tabs, sidebars, chat histories, upload buttons.
The chat window had a tax that every power user knows: copy something out of an existing chat or a document, switch tabs, paste it into a new or existing chat, explain all the context the AI cannot see, wait, copy the answer, switch back, paste it in, fix the formatting, and hope nothing got lost along the way. Do that dozens of times a day and the minutes add up to hours.
Notice who is doing all the actual hard work: you my dear human friend.
The AI just generates a response, but you are its carrier. You are it’s hands, body, instructor and slave all at once.
The third stage removes your labour.
Agents in the terminal, Claude Code being the one that broke the record: read your files themselves, run the commands themselves and write the results back where they belong.
Claude Code’s birth matches, if not exceeds, the ChatGPT moment of early 2023. Which is why a lot of people like to call it Claude Computer, because coding is only where it started.
And then came a fully family of competitors: Cursor, Codex, OpenClaw, Hermes, OpenCode, Kimmi CLI.
And more just keep coming. The main difference is what models they can use, but amongst every other feature they go neck to neck.
But then why do people still rave about Claude Code?
It gives you the cheapest access to a Claude Model with all the competitive features.
Using your Claude Subscription makes it 5x to 10x cheaper than using an API key directly. Kimmi CLI is the closest, underrated competitor because Kimmi’s K3 competes with Claude’s Fable 5.1.
So the progression is:
- AI models could only be reached and used by developers or by someone who knew how to call an API (not easily usable)
- A chat window you could reach as anyone, but the human did all the work (usable but unproductive)
- And now an agent that uses the computer itself (productive)
Why does an AI that can use the terminal matter?
Every click, every keystroke, every switch between screens you as a human do costs a few seconds. Across a day, that is a lot of seconds. Hours actually.
Think about converting a PDF into a Word document the way most people still do in 2026.
You Google “PDF to Word converter online”. You click a link. You click the upload button. You find the file in your folders, drag it in, press submit, wait while it is processed on someone else’s server, and download the result.
In the terminal, that is a one line command.
The terminal is a black window where a writer, a developer or an AI can run a command.
A command for what? To perform an every day operation:
| What you wanted | What it is in the terminal |
|---|---|
| A PDF turned into a Word document | pdf2docx convert report.pdf report.docx |
| A draft turned into a client-ready PDF | md-to-pdf proposal.md |
| Two PDFs merged into one | qpdf --empty --pages a.pdf b.pdf -- merged.pdf |
| The annoying iPhone HEIC photo as a JPG | magick photo.heic -resize 1600x -strip web.jpg |
| Stripping GPS data from photos before they go on social media | exiftool -gps:all= *.jpg |
| Pulling the text out of 40 PDFs so you can search them | pdftotext -layout *.pdf |
| Which of 4,000 documents mentions one person | rg "Jonas Smith" ~/Documents |
| Pulling the audio out of a video call | ffmpeg -i call.mp4 -vn audio.mp3 |
| A newsletter draft turned into a styled web page | pandoc post.md -s --css style.css -o post.html |
| A screenshot of any web page, for your post (Simon Willison’s tool) | shot-scraper https://example.com -o page.png |
| A web page saved so it still opens in five years | monolith https://example.com -o archive.html |
| A screen recording turned into a GIF for Substack | ffmpeg -i demo.mp4 -vf "fps=12,scale=600:-1" demo.gif |
| An animated explainer video, written as code | npx remotion render src/index.ts Explainer out.mp4 |
| A flowchart drawn from a few lines of text | mmdc -i flow.mmd -o flow.svg |
| A markdown file turned into slides | marp deck.md --pdf |
These commands look simple but underneath each, someone 10x smarter has written a package (all the logic as code) to get you from ‘I want this’ to ‘I have this’. It matters because one typed command can replace the whole click-and-wait routine you do. And reminder, you click and wait multiples times a day. Hundreds of times a week. Thousands of time a month.
We spent decades and a lot of venture capital building interfaces so that humans could do, by hand, what three lines in a terminal could always do. Honestly, it is a bit funny.
The best developers I have known are good at exactly two things: solving a new problem every day, and then doing absolutely nothing after.
They found the path with the fewest steps, automated their own work first, and moved on.
Over the years they wrote library after library and turned them into small packages anyone can install and run from the terminal. Compress an image. Turn those annoying HEIC photos from your iPhone into a JPEG. Convert a PDF. Transcribe a voice note.
You don’t need to know what happens inside. You give it the file and it gives you the result.
The catch was always human memory. To use those commands you had to remember them, and that kind of memory only comes from practice, the way muscle memory does.
That is what kept the terminal for developers, but now Claude Code (out of the box) can use terminal commands with 90 to 99% accuracy
It has read the documentation for the packages.
It knows what goes in, what comes out and which options each one takes.
For common jobs you don’t even need to give it instructions. You say what you want, it installs the package if it is missing, and it does the job.
So the fluency problem is gone. You describe the outcome, and the terminal does the work.
The people of 2030 are using the terminal in 2026. They are lazy in the best way, exactly like those developers
Now let me tell you who I think wins the next few years. I call them the people of 2030.
They solve a problem once, hand it to their agent, and then do nothing about it ever again. They don’t open an app to do a job. They say what they want, and the job is done on their own machine.
What a day in the life of a 2030 human looks like
The errands you do in a browser tab over 15 minutes are one line, one second commands for them:
- A draft turned into a client-ready PDF and emailed to the right person
- An invoice from Amazon or Apple found from the email they made when they were 18
- An HEIC iPhone photo resized, compressed and converted so the rest of the world can accept it
- Converting their bank statement and categorising the transactions across ‘Eating Out’, ‘Groceries’, ‘Utilities’ and what not
Notice what all of those have in common: nothing got uploaded. Every free website that does these jobs is, underneath, a hand-me-your-personal-documents service.
Then there are jobs most of us quietly give up on.
Not hard jobs, just jobs that came in hundreds. Pulling the text out of forty PDFs so you can search them. Finding which of four thousand documents mentions one client.
For the people of 2030, two hundred files cost the same effort as one.
That is not a time-saving claim. It moves functional data and information work to the AI and leaves the human with the strategic, creative work that requires thinking. Or just time to take a holiday really. Or have a fully automated business that runs itself when they’re not/no longer available.
Running Claude Code is light and a Macbook Air is just right
Fun fact: Your MacBook Air is idle about 95% of the time you use it. Open Activity Monitor in the middle of a busy day and watch the chip sleep between your keystrokes.
The AI model runs in the cloud, and your Mac only does the converting, searching and file handling.
By 2030, I think computers split into two camps: a light machine for people who keep the AI in the cloud, and a heavy one for people who run the AI on their own device to keep costs down. It’s either an Air or a Max.
The middle disappears, and clicking through apps becomes something we do very rarely.
Here’s why and how the terminal lets you easily build a paid Substack newsletter business
I didn’t write this newsletter on my laptop, I wrote it while driving. Well, I spoke it. Out loud.
Today (9th October 2026), I was on my way back from Qi Gong practice. Somewhere between the practice and home, the whole thing arrived at once: death, the business that has to survive it, and why the terminal is the answer. So I spoke it, right there in the car, in eighteen voice notes between 7:26 and 8:05 pm.
Now just like yours, my ideas show up at the most inconvenient times. So I HAVE TO capture them instantly
Capture is one press on my iPhone 17, the Action Button runs an Apple Shortcut. A Shortcut is a small automation you build on your iPhone without writing code, and this one does 3 things:
- It listens
- It transcribes what I say into text
- It saves the text as a markdown file in a folder on my iCloud Drive
Why a file and why not an Apple Note?
A markdown file is easy to read line by line for Claude Code, without overloading the context window or burning tokens.
Same goes for Codex, Kimmi CLI, Cursor, Hermes, OpenClaw, etc. And more importantly, it’s a file format that doesn’t require a specific software to be installed. It works on any device, because all it needs is the terminal or the note pad app.
iCloud then carries that file to every iPhone, iPad and Mac I own. By the time I sit down, it is already in my knowledge folder, ready for Claude Code to read.
The Knowledge Work Coach builds this capture button with you, step by step, and points it at a folder on your own iCloud.
That is the heaviest part, and it should be. Some of what you have read so far is not easy to say out loud. Letting the emotion out while I speak lets it show up in the writing, and it takes the emotion out of the editing when I get to it.
I can search years of my dications, published newsletters and drafts by meaning without compromising my privacy
Once your knowledge is in, what you need as a writer is a way to search your own knowledge by meaning, not by keyword.
Semantic search is the plain name for exactly that.
It takes every line in your notes, drafts and past newsletters and places them on a kind of map, where lines that mean similar things sit close together and unrelated lines sit far apart. When you search, you get back whatever sits closest to what you meant, even if you used different words.
Until recently this only ran in the cloud. Now it runs on a MacBook Air. A small but powerful search engine that keeps your notes on your own machine and lets Claude Code search them by meaning. The Knowledge Work Coach installs it and makes every folder of your notes searchable this way.
And it makes writing easier because we forget things
You and I forget. That is healthy.
But when your thinking lives in files and can be searched by meaning, forgetting stops costing you. It feels even healthier and less burdensome.
More importantly, it lets the AI arrive at the same point you are at. Your train of thought: what’s going through your head in that moment.
When Claude Code can search your own knowledge, it surfaces the exact memory you need, and you are working on the same wavelength. That is when it stops being an assistant and starts being a thinking partner, a co-founder even. (I refuse to say copilot. I just don’t like the word.)
Can you believe what I just did? I wrote about the terminal and Claude Code from the ‘if I die’ point of view.
This newsletter did not come out of Claude Code on its own.
It started from voice notes about building a business that is prepared for my short, medium or long term absence.
Searching my own ideas by meaning surfaced that thread, I read it back, and the angle came: the terminal as a safety net that keeps the business alive when I am gone.
That angle is the part AI cannot do.
It comes from me, a living, breathing, sensing and thinking human who just happens to note it down. What the search does is bring my past thinking to the surface fast enough for me to use it.
And here is the part that matters for writing a newsletter.
If my knowledge is there, searchable, growing every day, then when someone else sits down at my computer they can see the trains of thought that went through my head and pick up where I left off.
They can understand what decisions I made, when I made them and why I made them.
That is one part of the answer. Not the whole thing. So hold your horses.
Next, I keep up with the writers I follow using good old RSS
My own thinking is only half of what I search. The other half is what I read.
Every writer worth following publishes a feed, called RSS. It is one of the oldest things on the internet: a plain list of a person’s latest posts that any programme can read. Substack gives every newsletter one, at its address followed by /feed.
I keep one plain text file with a line per person. Right now it lists 31 writers, including Michael Crist, Teresa Torres, Ethan Mollick and Simon Willison. A small script reads every feed a few times a day, saves any post I don’t have yet as a markdown file, and gives each person their own folder. Adding someone new is one line in that file.
Nothing to host and no subscription. Everything they publish lands on my Mac, one collection per person, so searching one writer never mixes them up with another. The Knowledge Work Coach finds each writer’s feed, sets up the script and runs it on a schedule.
But I don’t just keep up, I can also search what they wrote by meaning, and link back to them
Then the same search by meaning runs over their posts as well as mine.
While I’m editing, if a line feels like it connects to something, I point Claude Code at that one sentence and tell it “here’s the scramble and babble of who/what I think it relates to”.
It searches across my dictations, my newsletters and everything the people I follow have published. What comes back is a handful of exact snippets, not whole articles. That keeps my head clear, and it keeps the AI’s context clear too.
I read those few lines and see who shaped the idea I’m writing about. Then I write the paragraph better, and I link back to them.
That is how this newsletter credits SemiAnalysis, SitePoint and Ethan Mollick in the section on how AI moved from the chat window to the terminal. I didn’t remember where I had read those ideas. The search did, and Claude Code put the links in, because hunting down exact URLs is a pain.
I read newsletters every day for exactly this reason. AI is making a lot of people dumber, and I don’t want to be dumb. Linking back is how I keep the reading honest and give the credit where it is due. Because the Knowledge Work Coach keeps your notes and their posts in the same search, this works from day one.
Editing and tracking: git
The last piece is git.
Git is, on the face of it, a tool developers use to keep track of code. The reason it matters for a writer is not the code part. It records every change to a file, line by line, with a date and a reason.
So when Claude Code edits a paragraph in this newsletter, I see a diff: the old line in red, the new line in green, side by side. No more wondering whether it quietly rewrote my intro. I can see exactly what changed, accept it or undo it. Before each round of edits, I save a snapshot, so I can always go back.
No more “final-v2”, “final-final-v3”. One file, and its whole history underneath it. The Knowledge Work Coach turns your knowledge folder into a git folder on day one and keeps it in the background, so you never have to learn the commands.
And go back to the question this newsletter started with. When someone else sits at my computer, git shows them not just what I wrote, but how every piece changed and when. My thinking, the writers who shaped it, and the history of every edit, all in one place.
So here’s the whole flow, from a voice note to a finished newsletter
- Speak it. One press of the Action Button, and the idea is a markdown file on my own iCloud.
- Shape it. Claude Code turns the ramble into headings and paragraphs, in my voice.
- Relate it to me. Search by meaning over my own dictations, drafts and newsletters.
- Relate it to the people I read. The same search over the writers I follow, synced through RSS, with links back to them.
- Edit it. By hand, top to bottom, with git keeping every change.
AI never does the thinking in any of it. It formats, it searches, and it puts the links in. The thinking, the angle and the edit stay mine.
This evening, that flow took me from eighteen voice notes in the car to the piece you are reading. And none of it needs me to be a developer.
The agent that sets all of this up for you is the Knowledge Work Coach. It walks you from opening the terminal for the first time to the capture button on your iPhone, search by meaning over your own notes, the writers you follow synced to your Mac, and git quietly in the background. About an hour, start to finish, and $36 once.
The 3 parts that make up an AI Agent and how to form a Second Brain Operating System
The rest of the answer is three words that get muddled constantly, so let me separate them.
Knowledge
Knowledge is what you know, stored so it can be searched by meaning. It can be huge, and it does not all need to be loaded at once. The AI pulls up only the few pieces that are relevant, answers from them, and keeps the rest out of the way so it doesn’t get overwhelmed.
The technical name for this is retrieval augmented generation. The search retrieves, the AI augments its answer with what it found, and then generates. That is all it means.
Skills
My knowledge is enough to decide what a featured image for a newsletter should say: the hook, the wording, the angle. It is not enough to decide how to put that image together well.
That is a skill. A skill is a best practice written down: the way to do a job properly. It is what a prompt used to be, with one big difference. A skill names the terminal commands to use.
Take a report. The content, the sentences, what page one says, come from your knowledge. The formatting, how the tables are laid out, how long it should be, whether it ends up as a PDF or a Word file, come from the skill.
Skills lean on terminal commands because commands give the same output for the same input, every time. When you hand your business to someone else, predictable is exactly what you want.
Tools
Finally, tools. Some jobs will always live on the internet: your email, your calendar, your website. To reach those, Claude Code uses MCP.
MCP, the Model Context Protocol, is the unglamorous name for a standard plug. Any online service can expose what it can do through MCP, and Claude Code can then use it the same way every time. Some people will tell you MCP is outdated. It isn’t. It is how the AI does the last step, like actually emailing the report to whoever needs it.
Put the three together and you have an agent
As promised, everything you need to run a paid Substack newsletter business that can handle your absence
1. The foundation: the Knowledge Work Coach. Everything in this letter so far. Capture from your phone, your notes and the writers you follow searchable by meaning, git in the background. Every other agent stands on this folder. It also opens your Second Brain OS account, which comes with an AI receptionist you can put on WhatsApp, Instagram, Telegram and your website.
2. Then it reads you back: the Productivity Coach. The Knowledge Work Coach writes, the Productivity Coach reads. Daily recaps of what you actually did, a weekly look back, and next week’s plan, with the tasks put straight into Todoist. It also lets you run all of this remotely, so you can use the Second Brain Operating System on your Mac from your iPad Mini or iPhone. It’s how I spun up Claude Code on my Mac from my phone to clean up my voice notes when I parked my car at home and had dinner with my wife instead of going to the computer.
3. Then your public face: a public clone that works as an AI Receptionist. Your readers message you. Someone has to answer, at 3am, in their language. The clone does, on WhatsApp, Instagram, Telegram and your website, and it answers from what you have already published. For a writer, that is years of newsletters, so a reader asking a question at midnight gets the answer you would have given.
4. Then it books the call: the full appointment lifecycle. It checks your real calendar and books the call inside the conversation, then finds, moves and cancels the bookings you already have.
5. Then the website that sells: the Website Operator. It takes a website from deployed to live, measured and selling. Your domain, your analytics, the clone (your AI Receptionist) on every page, and the design passes, run from your terminal without a developer. It is also how you finally see whether readers from Substack behave differently from the ones who arrive from Instagram.
Five agents, one Second Brain Operating System. Your published newsletters become your clone’s knowledge, your clone books the calls, and your website shows you which of your writing actually brings people in. None of it needs you awake.
Each one carries its own skills and tools. Each one can access your knowledge and you get to decide the limits of that access.
Together they run the parts of your business that don’t need you to be active all the time. You type one line, and the agent already knows what to search, how to do the work and where to send it. You stop explaining the job every time.
The Knowledge Work Coach is the first agent, and the rest are on agents.umairkamil.com. You pay once for an agent, it sets itself up on your Mac and it runs there, with no monthly bill. The few tools that can only live online, like managing your calendar, are pay as you go: you only pay for what you use.
And as always, I can setup everything for you, with you. All it takes is a call.
Why this is what survives you
Go back to the question at the start. Could someone sit at your computer and run things the way you do?
With a Second Brain Operating System, a lot more of the answer is yes. Your thinking is in files, searchable by meaning. Your best practices are written down as skills. The tools you use for emails, calendar and more are attached. The history of every change is in git. Whoever sits down next doesn’t have to guess how you worked. They can read it, search it and run it.
And if any of those parts is missing, it isn’t a Second Brain Operating System. It won’t stand the test of your death, whatever business you run.
I started this letter thinking about the day I die, and about the people who would wake up a week later to my absence. I can’t do much about the first part. The second part, I can.
My thinking is in files. The writers who shaped it are a search away, with links back to them. My best practices are written down, and every change is on record. Someone who loves me could sit at my computer and keep the ship moving.
And on the days I am very much alive, the same setup gives me something I never had while the business needed my hands on every lever. A day off.
Your death is also a holiday. Just a much longer one. Prepare for the worst, hope for the best. Build it for the first and take a holiday, worry free.
If you want to start, start with your knowledge. The Knowledge Work Coach gets your thinking out of your head and onto your own machine, searchable by meaning, in about an hour. $36.
P.S. One more piece of advice while I am here. Don’t build your business only on coaching or consulting. If it depends on you showing up, it dies with you. Have digital products. Have a paid newsletter. Have a mix of things that work without you. I wrote more about that in Optimise for the life, not the money.
P.P.S. If you’re down to get started with a Second Brain Operating System, here’s how
