Pillar essay · May 27, 2026

Why CEOs should use Claude Code in the terminal

From chat tabs that forget your business to a folder that knows it. Four primitives that turn Claude into the leverage layer of your week.

The year of the chat tab is over

A year ago, most CEOs had never heard of Anthropic or Claude. Today, pretty much all CEOs have a paid Claude account and open it every day.

They're still using it in a tab. Chat windows. Pasting context in, pasting outputs out. Starting over every Monday because Claude forgot everything that happened last Tuesday.

The chat window is cool. It's also like baby Claude.

The actual product, the one that turns Claude into leverage you can feel by the end of the week, is Claude Code running in a folder on your laptop. You open a terminal. You type claude. Claude reads the brief in that folder, knows what your business is, who your customers are, where your files live, and what tone you write in. You ask a question that would have taken three paragraphs of setup in a chat tab. The answer comes back in seconds, already in your voice, already referencing the right files.

That's the move. The next twelve months are going to separate the CEOs who made it from the CEOs who didn't, and the gap will be bigger than people expect.

This piece is the strategic case for making it. If you want top 1% leverage from AI as a CEO, learning Claude Code is the most important thing you can do this year. The good news is that the unlock is four small primitives, and you don't have to be a developer to install any of them. You just have to open a folder on your laptop.

Even if you've never opened a terminal in your life.

Your Claude actually knows your business

The thing that turns Claude from a generic AI into one that knows your business is a single text file in the folder where your work lives. It's called [CLAUDE.md][1]. It's plain markdown. You write it once. Claude reads it every time you start a session, and from that moment on, it shows up to the conversation already knowing the things you'd otherwise have to re-explain.

The CLAUDE.md in a folder on your laptop is the thing that makes Claude actually know your business.

I opened the Headphones.com folder on a Tuesday afternoon a couple of months ago to ask a real question. We were debating whether to keep a specific SKU live through the back half of Q2 or sunset it. The question had numbers attached, opinions attached, and a customer-feedback thread on Slack attached. In a chat tab, that question would have been a fifteen-minute paste-fest. Paste the SKU performance. Paste the margin. Paste the Slack thread. Paste the customer comments. Half the answer would be Claude re-asking me for context I'd already pasted.

In the folder, I asked the question cold. Claude already knew what category we're in, what our margin profile looks like, who our top three customers are by revenue, and which Slack channel is where the real signal lives. It pulled the numbers from the right place, read the right thread, and came back with a four-bullet recommendation in under a minute. The answer wasn't magic. The context was already there.

That context lives in the CLAUDE.md. Mine is forty-something lines. It names what Headphones.com does, what scale we're at, what files matter, what tone I write in, and what NOT to touch. I wrote it once, six months ago, and I've edited it maybe four times since. Every minute of writing has paid back a hundred times in conversations that started warm instead of cold.

Sure you can create projects or artifacts in the chat app and memory is starting to become a thing there too. The chat side keeps improving and you can get partial versions of what I'm describing without leaving the browser. But a terminal session opens in a folder on your laptop, reads CLAUDE.md, and knows exactly where to look for ALL the context that's relevant to your business. Even better, if you set things up properly, it picks up the context your business creates automatically: new Granola transcripts, new commits, new notes, new decisions. The value compounds every day.

If you want to go a level deeper on what the file is and what a good one looks like, the [explainer is here][2]. If you want the thirty-minute setup, the [how-to is here][3]. Both pair with this piece naturally. The point of this section isn't the file itself; it's that the file is the difference between Claude that re-asks you the same five context questions every conversation and Claude that opens with the answer.

Claude does the work, not just the suggestions

The second primitive is the one that breaks the chat-window mental model the hardest. In a chat tab, you describe what you want, the AI suggests, and you do. You're the actor; Claude is the consultant. In the terminal, that flips. You describe the outcome, and Claude reads files, writes files, runs commands, calls connectors, and reports back. Claude is the actor; you're the orchestrator.

In the terminal, the agent is the actor and the whole mental model of using AI shifts.

The cleanest illustration I run is on top of [Granola][4]. Granola records every meeting I sit in. The transcripts get synced to a folder of markdown files on my laptop, one per meeting, organized by date. That whole [setup is its own walk-through][5] and it's the prerequisite for almost everything I do on top of it.

Once the meetings folder is live, I have a skill that triages every new transcript end-to-end. I finish a 1:1. I drop the transcript into the folder. Claude reads it, extracts the decisions we made, the commitments I owe people, the asks the other person is waiting on me for, and the customer truths worth saving for content. The decisions get appended to a running decision log. The commitments get added to a commitment ledger that's already organized by person and date. The asks land in my outbox as draft replies in my voice, ready for me to skim and send. The content-worthy moments get filed into an ideas backlog.

In the time it takes me to walk from the meeting room to the kitchen for water, the agent has done thirty minutes of work I used to do by hand. I didn't write the code that runs the triage. I wrote a brief once, six months ago, that tells Claude what to do with new transcripts. Claude has been doing it ever since.

The mental shift is the part most CEOs miss when they hear "agentic." Agentic doesn't mean smarter AI. It means a different verb. In chat, you say: "Help me draft three follow-ups based on yesterday's call." In the terminal, you say: "Draft three follow-ups based on yesterday's call and save them to my outbox." Then you go get coffee. When you come back, the three drafts are there. You skim, you edit one line, you send. The work happened while you were getting coffee.

You can compose this with [Routines][6], which run agents on a schedule. The triage skill runs whenever a new transcript lands. A weekly review skill runs every Friday at 3pm. A morning brief runs at 6:30am. The agent shows up to you, not the other way around. Combine that with the persistent memory from the section above and you have a stack where the AI is doing real work, with full context, on its own clock.

This is the part of Claude Code that genuinely doesn't have a chat-window equivalent. Projects and artifacts can hold context. They cannot, in May 2026, read forty markdown files in a folder, edit eight of them, write two new ones, and run a script that updates a fifth. They live in a browser tab. The terminal lives on your computer, with access to your files, your shell, your scheduled jobs, and the connectors that talk to the rest of your stack. That access is the whole game.

Skills are how you stop repeating yourself

A [skill][7] is the unit of compounding leverage in Claude Code. The technical definition is small: a skill is a folder with a markdown file in it that tells Claude what the skill does and when to invoke it. The strategic definition is bigger: a skill is how you turn a workflow you ran once into a capability Claude executes on command, forever.

A skill turns a workflow you ran once into a capability that runs forever.

The skill I always come back to when I'm walking a CEO through this is one I call /learn. It's tiny. It does one thing. When I finish co-editing a piece of writing with Claude, I run /learn. The skill reads the diff between the AI's draft and my final shipped version, extracts the principles I applied in my edits (often principles I was applying without being able to name), formats them in the shape my voice profile uses, and appends them to the profile.

I wrote the skill in about forty minutes, six months ago. Since then it has bumped my voice profile from v0 to v10. Every new AI draft uses the current version. The drafts I get back today are noticeably closer to publishable than the drafts I got six months ago, because the system that writes them has six months of my edits already loaded.

The skill is the unit that makes that compounding possible. Without the skill, my edits would have been useful exactly once, in the moment, and gone. With the skill, every edit teaches the next draft. The taste compounds.

I have a stack of skills like this now. Some are mine; some I pulled from the open community. A skill that triages my inbox. A skill that drafts pre-meeting briefs. A skill that wraps the Granola transcript triage I described above. A skill that builds a weekly review every Friday. Each one took an evening to write. Each one runs forever. The week-over-week effect is that I'm prompting Claude from scratch less and calling capabilities I already built more.

If you've heard the term [slash command][8] used in Claude Code conversations, it's the same idea. A slash command IS a skill, packaged so you can invoke it by typing /skill-name. Same idea, lighter packaging. The deeper [explainer is here][7] and the [thirty-minute walk-through to build your first one is here][9]. The reason both pieces exist is that skills are the part of Claude Code where the chat-window CEO and the terminal CEO start to look like they're running fundamentally different operations. The chat CEO is re-describing the workflow every Monday. The terminal CEO is calling the workflow by name.

Skills are also portable. The open agentskills.io standard means a skill folder you write in Claude Code today also works in the open-source harnesses above it ([OpenCLAW][10] and [Hermes][11]). The capabilities you build don't lock you into one tool. That portability is most of why the skills layer is where I tell CEOs to invest first.

It compounds every week

Memory, agentic execution, and skills are the three Claude Code primitives. The fourth primitive isn't a feature; it's the shape of the thing those three produce when you let them run for three or six or twelve months. The week-one effect of installing Claude Code is real but modest. The month-six effect is generational.

Every week of running Claude Code in your folder compounds. Every week of running Claude in tabs resets.

The compounding is the part I struggle to communicate to chat-window CEOs because the math is invisible from the outside. From the outside, it looks like one CEO has a paid Claude account and another CEO has a paid Claude account, and they're both "using AI." From the inside, one of them is running a stack where the memory deepens every week, the skills library grows every month, and the agent's grasp of the business gets sharper against every edit. The other is starting from zero every Monday.

In eighteen months, the gap between those two CEOs will be larger than the gap between either of them and a CEO who never installed AI at all. Neither will be working harder than the other. The compounding is the difference.

This is also the altitude where the [harness][12] layer above Claude Code starts to matter. A harness is the orchestration layer that wires Claude Code to the rest of your stack: connectors to your CRM and inbox and calendar, persistent memory across sessions, scheduled routines, multi-agent delegation. The two open-source harnesses worth knowing about today are OpenCLAW and Hermes.

[OpenCLAW][10] is the harness I run on Headphones.com and Lantern.is. It's open-source, community-stewarded by a non-profit foundation, and the most mature ecosystem in the space. The community library of connectors and skills is large; the operator's manual exists (I wrote the 270-page one); the install is laptop-resident, which matches how most CEOs already work. If you want the deeper walk on what OpenCLAW is and why I run it, the [explainer is here][10].

[Hermes][11] is Nous Research's open-source harness, the most visible alternative to OpenCLAW in the open ecosystem. It runs standalone, leads with a Telegram-first mobile interface, and works with multiple models including Grok via X Premium credits. I haven't switched. The operators I respect who run Hermes describe a meaningfully different operating model, especially CEOs who run their week from their phone instead of their laptop. The [explainer on Hermes is here][11]; the [side-by-side comparison of the two is here][13] if you're deciding between them.

Whichever harness you eventually run, the skills you write today work on both. The portability is real, and it means the work you put into Claude Code at the foundation layer compounds across whichever harness you end up choosing.

The compounding isn't an opinion. It's the math of doing the same work in two different shapes. In the chat tab, every conversation is fresh; the value of last month's work is whatever you remembered to copy out. In the folder, every conversation builds on the last one; the value of last month's work is in the memory, the skills, the routines, and the harness on top. The first shape adds up linearly. The second shape compounds.

Do this now

The whole pitch of this piece is that the terminal isn't technical because of the keystrokes. It's technical because of the leverage. The keystrokes themselves are the easy part. The hard part is deciding to leave the chat tab.

The single move I'd ask you to make this week is to download the sample chapter of The Complete Guide to Claude Code for CEOs. It's free. It walks the install end-to-end, including the part most CEOs are nervous about (opening a terminal for the first time), and it ends with you actually running Claude in a folder on your laptop.

Get the sample → [desktheory.com/sample/claude-code][14]

The full book is on the same page when you're ready. It takes you from no Claude Code install at all to five end-to-end workflows running on your laptop against your own files. The promise is what you've read above: by the time you're done, you'll be comfortable building high-leverage solutions to issues across your business, and you'll be doing it in the terminal.

Even if you've never touched the terminal in your life.

Tell me in thirty days what changed. I love hearing from CEOs who made the jump.

Andrew

Why CEOs should use Claude Code in the terminal

The year of the chat tab is over

Your Claude actually knows your business

Claude does the work, not just the suggestions

Skills are how you stop repeating yourself

It compounds every week

Do this now

Related reading

The architecture behind these articles.

Why CEOs should use Claude Code in the terminal

The year of the chat tab is over

Your Claude actually knows your business

Claude does the work, not just the suggestions

Skills are how you stop repeating yourself

It compounds every week

Do this now

Related reading

The signal in your inbox, every Thursday

The architecture behind these articles.