Skip to main content

The Minimum Viable Jarvis

A step-by-step guide to setting up the simplest system that gives you AI-augmented recall, strategic clarity, and compounding context.

Time estimate: 3.5 to 4 hours to complete everything in this guide, even with some prior technical experience. If you are completely new to the terminal and have never installed developer tools before, expect the upper end.

The value of an instructor: This guide is designed to be self-paced, but having someone experienced walk you through it in person makes a significant difference. Every machine is slightly different. You will hit edge cases (a Windows PowerShell permission error, an old Node.js version conflict, a corporate firewall blocking a download) that are too niche to document here but take 30 seconds for an experienced person to debug. An instructor gets you across the finish line instead of stuck at Step 1B for an hour. If you have access to an Applied AI Society workshop or know someone who has done this before, take advantage of that. If not, this guide plus Claude Code itself (which can help you debug installation issues) will get you there.

Why This Matters​

You have a dozen inboxes. Discord, Telegram, iMessage, email, Slack, LinkedIn, X, Instagram, phone calls, in-person conversations. Right now, you probably have 50 open threads across 12 platforms. No human brain can track all of that. And the honest truth is: you're dropping balls. We all are.

The thing nobody tells you about leveling up as a professional or leader is that the job changes underneath you. At a certain point, the most important work is no longer doing the work. It's defining reality, setting objectives, and evaluating whether the system is working. You shift from working in the business to working on the business. Meta work becomes the work. This shift from execution to design is what we call game design: the discipline of defining objectives, rules, guardrails, and scoring for the AI agents in your system.

Here is the uncomfortable truth: you are the bottleneck. Not the tools. Not the AI. You. The quality of your strategic thinking, the clarity of your communication, and your willingness to document what you actually know are the limiting factors. That is not a criticism. It is empowering. Because if you are the bottleneck, you are also the one who can unblock everything. And AI can help you see your own thinking more clearly, pressure-test your strategy, and refine your plans in ways that used to require an expensive advisor or a very patient co-founder.

Meta thinking is the new thinking. The highest-leverage skill you can develop right now is not execution. It is the ability to design your business as a system: the objectives, the rules, the guardrails, the scoring. Execution is increasingly commoditized. Your ability to define what should be executed is not.

Here's the key insight behind everything that follows: the truth in your head is not the truth. Not operationally. Not for AI. Not for your team. The truth that matters is the truth that exists in documents that AI can read and act on. If it's only in your head, it might as well not exist. It's ungreppable. It's unsearchable. It's locked in wet storage with a terrible API.

Your Sovereign Agentic Business OS is the persistent memory your AI draws on. The Minimum Viable Jarvis (MVJ) is the simplest possible version of that business OS. Not the end state. The starting point.

For the full philosophy behind why documented truth matters, see Truth Management and Why It Matters. This playbook is the practical "how to start" companion to those ideas.

Phase 1: Install Your Tools​

Most of what you need is free or cheap. The entire stack can be running in under 30 minutes.

Before you start: if you are unfamiliar with any of these tools and want to verify they are safe and legitimate, that is smart. You can paste the link to this tutorial into any AI chat (ChatGPT, Claude, Gemini) and ask: "Is this all safe to install? What does each tool cost? What are the advantages?" It will walk you through every tool listed here. Spoiler: almost everything is free and open source. The one paid tool is Claude itself (~$100/mo for heavy usage).

Here is the cost breakdown upfront:

ToolCostNotes
Node.js / npmFree, open sourceRequired to install Claude Code
Claude Code CLIFree to installRequires a Claude account. Free tier is limited. Claude Max (~$100/mo) for daily usage.
VS CodeFree, open sourceMade by Microsoft
GitFree, open sourceVersion control
GitHubFreePaid tiers exist but you do not need them
SuperwhisperFree tier availableVoice-to-text, fully local
Wispr Flow~$10/moVoice-to-text, cloud-based
GranolaFree tier availableMeeting transcription (optional)

Step 1A: Voice-to-Text​

The bottleneck between thought and text must be removed. This is not just about speed (though speaking is 3 to 5x faster than typing). It is about flow states.

When you are typing, part of your brain is thinking about typing. You are compressing what you would otherwise say because the friction of getting it out is too high. You edit yourself mid-thought. You lose threads. You stay in the analytical, word-by-word part of your brain instead of the big-picture, strategic part.

When you speak, you stay in flow. Your brain operates at its best capacity. Ideas connect to other ideas. Two hours fly by and you realize you just produced a massive amount of high-quality thinking. That is the state you want to be in when you are working with your Jarvis.

Two solid options:

  • Superwhisper: Fully local, privacy-focused. Your audio never leaves your machine. Great if sovereignty matters to you (and it should).
  • Wispr Flow (~$10/mo): System-wide dictation that works across any application. Slightly more polished UX. One great feature: it auto-reformats what you say. If you stumble, say "oh wait," or restart a sentence, it cleans all of that up. It adds line breaks and structure to your raw speech. The output is surprisingly clean.

Either works. You hold a key, you talk, you release, and the text appears wherever your cursor is. It works with every application that has a text input: your terminal, a browser, Slack, email, a Google Doc, anything. Wherever your mouse clicks into a text box, that is where the transcription goes. Wispr Flow also keeps a history of everything you have dictated, so you can go back and copy-paste a previous dictation into a different app if needed. Once you install it and start using it, it works everywhere. Even whispering works, which matters if you are in a co-working space or a meeting. The point is that you can speak naturally and get text. For more on the role of voice transcription in truth management, see Voice Transcriber.

Step 1B: Install Node.js and npm​

Claude Code is installed through npm (Node Package Manager), which comes bundled with Node.js. If you have never installed anything from the command line before, you need Node.js first.

If you try to run the Claude Code install command and see npm: command not found, this is why. You need Node.js installed first.

macOS:

The easiest option is to download the installer from https://nodejs.org. Run the .pkg file and follow the prompts. Done.

Alternatively, if you prefer using the terminal, you can install via Homebrew. If you do not have Homebrew yet, install it first by opening Terminal (search "Terminal" in Spotlight or find it in Applications > Utilities) and pasting this command:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Follow the prompts. Once Homebrew is installed, install Node.js:

brew install node

Windows:

First, find your terminal. Search "Command Prompt" or "PowerShell" in the Start menu. Both come pre-installed on every Windows 10 and 11 machine. On Windows 11, you may also have "Windows Terminal" which is a nicer wrapper around both. Any of these will work.

Then install Node.js:

  1. Go to https://nodejs.org
  2. Download the LTS (Long Term Support) version
  3. Run the installer. Accept the defaults. It will install both Node.js and npm.
  4. Important: Close and reopen your terminal after installing. The terminal needs to restart to recognize the new commands.

Verify it worked by typing this in your terminal:

node --version
npm --version

If both commands print a version number, you are ready to install Claude Code.

Step 1C: Install Claude Code​

Your terminal-based AI interface. This is the engine of the Jarvis. Important distinction: Claude Code on its own is not the Jarvis. The Jarvis is the combination of your file structure, your documented context, and how you use Claude Code to operate on all of it. Claude Code is simply the best engine for a Jarvis right now, but your files are yours. You could switch to a different AI tool tomorrow and keep everything.

Why not just use ChatGPT, Gemini, or any other chatbot? The big AI platforms (OpenAI, Google, etc.) want you locked into their ecosystem. Your conversation history lives on their servers. Your context resets every session or is trapped behind their interface. You cannot export it, version-control it, or run a different AI on top of it. They are designed to keep you coming back to their app, not to give you a system you own.

The MVJ approach is the opposite. Your files live on your computer. They are plain markdown. Any AI tool can read them. If Anthropic triples their price tomorrow, or a better tool comes out next month, you take your files and walk. No export process, no migration headache. You are not a user of someone else's platform. You are the operator of your own system.

This tutorial is not an Anthropic ad. We default to Claude Code because it is currently the best harness for this workflow. But the entire point of the MVJ architecture is that you are not locked in. You have options at every level:

  • Different harnesses. Claude Code is one harness. Others exist and new ones are emerging constantly: OpenCode, Cursor, Pi, Hermes, and more. They all read files and run commands. The instructions in this tutorial default to Claude Code, but the overall usage patterns (brain dumps, user profiles, skill files, relationship files) work with any harness that can read your workspace. Check the AI Dev Tool Power Rankings or Best AI Coding Agents comparison for current rankings.
  • Different models within the same harness. Even within Claude Code, you can swap models. You are not constrained to Anthropic's models. Other harnesses let you use any model you want.
  • Open source models you run yourself. Open source AI models are getting remarkably good, remarkably fast, and many of them can be run entirely on your own hardware with zero data leaving your machine. New ones come out every few weeks. Check the Open LLM Leaderboard or Open WebUI Leaderboard for current model rankings.

The competitive pressure between American companies, Chinese labs, and the open source community is driving quality up and cost down at a pace that benefits you. Today's best default is Claude Code. Tomorrow it might be something else. Your files do not care. That is what sovereignty means in this context, and it is worth thinking about as you decide which tools earn your trust and your money. For a deeper dive on sovereignty principles, see Sovereign Agentic Business OS.

npm install -g @anthropic-ai/claude-code

You will want a Claude Max subscription (~$100/mo) for meaningful daily usage.

Why the CLI and not the desktop app? Claude has a desktop app with a nice window, but it is a wrapper around the CLI, and the CLI updates far faster. Claude Code's CLI might update three times in a single session. The desktop app does not keep up, which means it is often buggy or missing features. The CLI auto-updates and is always current. Use the CLI in your terminal.

First launch: When you type claude for the first time, it will walk you through a setup flow. Here is what to expect:

  1. Theme selection. It will ask you to pick a color theme. If you do not have a strong preference, pick dark mode. It looks better in most terminals and is easier on your eyes for long sessions.
  2. Terminal setup. It will ask if you want to use Claude Code's recommended terminal settings (Option+Enter for newlines, visual bell). Say yes. Pick option 1: "Yes, use recommended settings." This makes the experience smoother. You can always change it later with /terminal-setup.
  3. Login. It will ask you to log in with your Anthropic account. Follow the prompts to authenticate.

Step 1D: Set Up the clauded Alias​

By default, Claude Code asks your permission before it does anything: creating a file, editing a file, running a command. Every. Single. Time. This is the safe, responsible design. And it will drive you absolutely crazy if you are trying to get real work done.

In practice, nearly every developer and power user runs Claude Code with the --dangerously-skip-permissions flag, which tells it to just do things without asking. The tradeoff is real: you are giving it permission to modify your files, run commands, and make changes without a confirmation step. But the productivity difference is enormous. With the flag, you stay in flow. Without it, you are clicking "approve" dozens of times per session.

The recommended approach is to create an alias so you have both options. The fun part: you can use Claude Code itself to set this up for you. Start a Claude Code session by typing claude in your terminal, then tell it:

"Add an alias called clauded to my shell config that runs claude with the --dangerously-skip-permissions flag."

Claude Code will find your shell config file (~/.zshrc on Mac, ~/.bashrc on Linux) and add the alias for you. It will ask for permission to edit the file. Approve it.

Once it confirms the alias is added, you need to do three things in order:

  1. Quit your Claude Code session. Type /exit or press Ctrl+C to get back to your normal terminal prompt.
  2. Close your terminal tab or window entirely. Your terminal loads the shell config file when it first opens, so the new alias will not exist in any terminal that was already open.
  3. Open a new terminal tab or window. Now the alias is loaded. You can verify by typing clauded and it should start Claude Code in dangerous mode.

If you do not want to use Claude Code to do this, you can manually add this line to your shell config file:

# Add this to your shell config (~/.zshrc on Mac, ~/.bashrc on Linux)
alias clauded="claude --dangerously-skip-permissions"

Either way, once the alias is set up and you have opened a fresh terminal, you now have two commands:

  • claude for when you want the safe mode with permission prompts
  • clauded for when you trust the operation and want to stay in flow

When to use which:

  • Use clauded for your day-to-day work: brain dumps, document creation, strategic thinking, updating your business OS. This is 95% of the time.
  • Use claude (safe mode) when you are doing something sensitive, like working with credentials, making changes to production systems, or any operation where you want to review each step before it happens.

Claude Code is not a chatbot. It reads your files, writes your files, runs commands, and operates within the context of your entire workspace. It is the brain that draws on everything in your business OS.

Phase 2: Set Up Your Jarvis Workspace​

Step 2A: VS Code​

Visual Studio Code is your window into the file system. Download it for free from https://code.visualstudio.com.

If you have never used a terminal before: the terminal is the text-based interface to your computer. When you see windows and icons on your screen, that is a graphical layer on top of what is really happening, which is your computer sitting in a folder, ready to execute actions. The terminal gives you direct access to that. You don't need to be fluent. You just need to be willing to open it. You will open the terminal inside VS Code in Step 2C.

Step 2B: Git and GitHub​

What is the difference between Git and GitHub? Git is a tool that runs on your computer. It tracks changes to your files over time, like an infinite undo history that also records what changed, when, and why. GitHub is a website (github.com) where you can store a copy of your Git-tracked files in the cloud, so they are backed up and accessible from anywhere. Think of Git as the engine and GitHub as the garage where you park your car. You need Git. GitHub is strongly recommended but technically optional. For a deeper explanation, GitHub's own guide is excellent.

Installing Git:

Setting up GitHub:

  1. Create a free account at https://github.com if you do not have one.
  2. Install the GitHub CLI so you can interact with GitHub from your terminal: npm install -g gh (you already have npm from Phase 1).
  3. Log in by typing gh auth login in your terminal and following the prompts.

Step 2C: Clone and Open Your Jarvis Workspace​

We have created a starter repo with the default folder structure for your Jarvis. You are going to use Claude Code to clone it to your computer. This is a good first rep of telling Claude Code to do something for you.

Use Claude Code to clone the repo:

  1. Open any terminal on your computer (you do not need to be in any particular folder).
  2. Type clauded to start a Claude Code session (using the alias you set up in Step 1D).
  3. Tell it something like:

"Clone the repo at github.com/Applied-AI-Society/minimum-viable-jarvis into a folder that makes sense on my computer for storing projects. If I don't have a folder for that yet, create one."

Claude Code will figure out the right location for your operating system. On Mac, it might put it in ~/Documents/github-repos/ or ~/Projects/. On Windows, it might use C:\Users\YourName\Documents\. It will ask you to approve the commands. Let it do its thing.

  1. Once the clone is done, note the folder path that Claude Code tells you it cloned into. You will need this in a moment.
  2. Type /exit to quit the Claude Code session.

Open the workspace in VS Code:

  1. Open VS Code.
  2. Go to File > Open Folder (or Ctrl+K Ctrl+O on Windows, Cmd+O on Mac).
  3. Navigate to the folder that Claude Code just cloned and select it.
  4. Now open the terminal inside VS Code. This is important: you want the terminal to be scoped to your Jarvis folder.
    • Mac: Terminal > New Terminal from the menu bar, or press Ctrl+`
    • Windows: Terminal > New Terminal from the menu bar, or press Ctrl+`
  5. In the VS Code terminal, type clauded. Claude Code is now running inside your Jarvis workspace and can see all the starter files.

You are now in the cockpit. The left panel shows your file tree. The bottom panel is Claude Code in your terminal. The right panel is for viewing whatever file you are working on.

If you do not have Git installed or prefer to start from scratch, you can skip the clone and tell Claude Code to create the folders for you instead. Start a Claude Code session and say:

"Create a folder called my-jarvis in a good location on my computer with subfolders for user, people, artifacts, meeting-transcripts, and skills. Also create a CLAUDE.md file that instructs you on how to operate as my business OS agent."

The starter repo comes with five folders:

  • user/ for your profile, voice, and anything that helps your Jarvis understand who you are
  • people/ for relationship files (one per person)
  • artifacts/ for strategic documents, decision records, status updates, and plans
  • meeting-transcripts/ for raw or processed transcripts from conversations
  • skills/ for SOPs that define repeatable tasks for your AI agent

It also includes a CLAUDE.md file that gives Claude Code instructions on how to operate within your workspace (this is what makes Claude Code understand the structure of your business OS from the first session), and a skill file that will interview you on your first session to create your user/USER.md profile.

Step 2D: Meeting Transcription (Optional)​

Tools like Granola run in the background during meetings and give you a transcript afterward. This becomes raw material for your business OS. Not every meeting needs to be transcribed, but the important ones should be captured so you can extract insights and commitments later.

Phase 3: Understand What Makes This Work​

Step 3A: The Chief of Staff Mental Model​

Think of your AI agent as a chief of staff. What does a chief of staff need to be genuinely helpful?

Tools. When your chief of staff can swipe your credit card, that is like giving your agent a tool. When they can access your calendar, that is a tool. Without tools, your agent is just a conversationalist. With tools, it can actually get things done: send emails, schedule meetings, look things up, run scripts.

Context about you. Your goals. Your decision-making style. Your risk tolerance. Your priorities. Your relationships and who matters in your network. The more your agent knows about the most important things in your life and business, the more it can act on your behalf without you having to correct it constantly. Without this context, you are just screaming into the void.

Standard operating procedures. Humans have SOPs. Agents have skill files: markdown documents that clearly describe exactly what the agent should do for a given task. Step by step, in plain English, often mixed with specific commands or scripts to run. You can co-write these with the agent (it knows how to talk to itself). Over time, your library of skill files turns your agent from a general-purpose assistant into a specialist that knows your operation.

A real example: The Applied AI Society is, as of March 2026, operated by one person. One. That one person runs events, writes newsletters, manages partnerships, creates strategic documents, drafts social media posts across platforms, processes meeting transcripts, maintains a CRM of hundreds of relationships, and publishes documentation. The way this is possible is a Jarvis with deep context.

For example, there is a skill file called aas-social-post that drafts social media posts for X and LinkedIn. When it runs, it does not produce generic AI content. It has access to every past post, every brand guideline, every event recap, every strategy document, and every relationship file in the system. So it knows what the organization sounds like, what has already been posted, what is being promoted right now, and who to reference. The output sounds like it was written by the person running the org, because the agent has enough context to actually represent them. That is the difference between a chatbot and a Jarvis. Context compounds. One person plus a deeply contextualized AI chief of staff can do what used to require a team.

The Minimum Viable Jarvis is about setting up the context layer: getting the truth about your operation, your relationships, and your thinking into files that AI can read. The tools and skill files come later as you grow the system.

Step 3B: A Note on Security​

As you connect more tools to your agent (email, calendar, file systems, payment processors), the surface area for things going wrong increases. This is worth being thoughtful about.

The principle is simple: human in the loop for anything consequential. Your agent can draft every email, but a human reviews before sending. It can prepare financial reports, but a human approves before money moves. It can suggest meeting responses, but a human confirms before commitments are made.

The risk is not that AI is malicious. The risk is that it is confidently wrong, or that someone finds a way to inject instructions into content your agent processes (a technique called prompt injection). If your agent is reading emails and acting on them without oversight, a carefully crafted email could theoretically trick it into doing something you did not intend.

Start with read-only connections and work your way up. Connect your calendar so the agent can see your schedule before you give it permission to modify it. Let it read your email before you let it send on your behalf. Build trust incrementally, the same way you would with a new hire. The MVJ as described in this guide is inherently safe: it is just files on your computer. The security considerations become more important as you expand into connected tools and automated workflows.

Phase 4: Build Your First Business OS​

This is what a first session looks like. Five exercises, about an hour total. By the end, you will have a working business OS with real data in it, a clear picture of your top strategic blocker, and an actionable plan for getting unblocked.

Step 4A: "Who Am I?" (15 minutes)​

This is the most important first step. Before your Jarvis can help you with anything, it needs to know who you are. The starter repo includes a skill file that will guide Claude Code through an interview with you.

Tell Claude Code:

"Read the skill file at skills/create-user-profile.md and run it."

Claude Code will interview you one question at a time. It will ask about who you are, what you care about, how you make decisions, the current state of your operation, and your biggest strategic blocker. Use voice-to-text. Speak naturally. Do not overthink your answers.

At the end, Claude Code will save a user/USER.md file that captures everything. This file is the foundation of your Jarvis. Every future conversation will be informed by it.

If you get stuck on a question and do not know the answer, just ask Claude Code: "Based on what you already know about me, what do you think?" It will offer its best guess, and you confirm or correct. This often surfaces insights you would not have articulated on your own.

The moment: Your Jarvis now knows who you are. Not the LinkedIn version. The real version. Your goals, your values, your decision-making style, and the thing that is actually blocking you right now. This alone makes every future interaction 10x more useful.

The user/ folder is not limited to USER.md. You can add any file that helps your Jarvis understand you better. For example, a user/voice-profile.md that captures your writing style, your tone, how you handle conflict, how you communicate with different audiences. That way, anything your Jarvis writes on your behalf actually sounds like you. The principle is: the whole folder is about the agent getting to know who you are. Customize it to whatever matters for your situation.

Step 4B: "What's My Plan?" (15 minutes)​

At the end of the user profile interview, Claude Code will ask if you want to think through your strategic blocker. Say yes.

This is the payoff. Claude Code now has deep context on who you are, what you are working on, and what is in the way. It will help you break your blocker into actionable steps and save the result as a strategic plan in artifacts/.

If you skipped this during the interview, you can trigger it yourself:

"Based on my user profile, help me create a strategic plan for getting past my biggest blocker. Save it as an artifact."

The moment: You walked in with a vague sense of being stuck. You are walking out with a concrete, written plan. Not generic advice from an AI that does not know you. A plan built from your actual context, your actual constraints, and your actual goals. This is what a Jarvis does.

Step 4C: "Who Do I Know?" (10 minutes)​

Create 3 to 5 relationship files for key people in your professional life. Use voice-to-text to dictate into Claude Code. For each person, capture:

  • Name and role
  • How you met
  • What you're working on together (if anything)
  • Last meaningful interaction
  • Anything you want to remember (their kid's name, that project they mentioned, the thing they're excited about)

Tell Claude Code to create a file for each person in the people/ directory (already set up in the starter repo from Step 2C). The format doesn't matter much right now. What matters is that these people now exist in your system.

The moment: These people now exist in structured form that AI can reference. You will never forget a detail about them again. The next time you have a meeting with one of them, your business OS can brief you on everything you know.

Step 4D: "What Did I Actually Decide?" (10 minutes)​

Create a strategic document capturing one major decision you've made recently. Dictate the story:

  • The situation: What was the context?
  • The options: What were you choosing between?
  • The discernment process: How did you think it through? Who did you consult? What factors mattered most?
  • The decision: What did you decide?
  • The aftermath: How did it play out? Would you make the same call again?

The moment: When your team (or future you) asks "why did we do it this way?", you point them here. The truth is managed. No more relitigating settled decisions from memory.

Step 4E: "My System Talks Back" (10 minutes)​

Now ask Claude Code to generate a briefing from everything you've created. Something like:

"Based on everything in this workspace, give me a briefing. Who am I? What's my strategic plan? Who are my key relationships? What decisions have I made? What should I be paying attention to?"

Watch what comes back. It won't be perfect. But it will be useful. And it will be drawn from your truth, not from generic training data.

The moment: Imagine this briefing after 30 days of adding to your business OS. After 90 days. After a year. Every conversation, every decision, every relationship, compounding into an increasingly rich and useful context. That's the trajectory you just started.

Phase 5: The Daily Workflow​

Once your MVJ is set up, the default interaction pattern is simple: you speak, the system listens and routes.

Step 5A: Open Your Workspace​

Open your business OS folder in VS Code. Open the terminal within VS Code (Terminal > New Terminal). Type clauded to start Claude Code.

Step 5B: Brain Dump​

Start talking. Voice-to-text into the terminal. Just dump whatever is on your mind. It might be a meeting debrief, a strategic thought, an update on a relationship, a new idea, a decision you need to make. Do not worry about structure. Just say what is true.

Step 5C: Let Claude Code Route It​

Based on what you said, Claude Code determines which existing documents to update, whether new documents need to be created, and how to maintain coherence across everything.

Step 5D: Review the Changes​

Look at what Claude Code did. Approve, correct, or refine. This is you being the "dictator of truth" for your operation. The AI proposes; you approve.

Step 5E: Repeat​

Over time, Claude Code learns the structure of your business OS and keeps everything consistent. Cross-references stay accurate. Outdated information gets flagged.

The brain dump is the lowest-friction way to keep your business OS current. You don't need to think about where information goes. You just need to say what's true, and the system handles the rest.

Phase 6: Common Pitfalls (From Real Sessions)​

These are real issues that come up when people set up their MVJ for the first time. Knowing about them in advance saves frustration.

Old computers will struggle. If your laptop is 8 to 10+ years old, expect installations to take longer, and some tools may have compatibility issues that nobody on the development team is prioritizing. This is just the practical reality of how software companies allocate engineering resources. Everything in this guide will work on Windows, Mac, and Linux, but if your machine is very old, consider upgrading when you can. The MVJ does not require a powerful computer, but a reasonably modern one (last 5 years or so) will save you a lot of frustration.

Slow internet makes tool installation painful. Claude Code downloads a binary during installation. If your connection is spotty, this can time out or hang. Make sure you have a solid connection for the initial setup.

Claude Code sometimes times out on long operations. This is normal. Just resume the session. Your files are already saved. Nothing is lost.

The multiple-choice UI in Claude Code can be confusing. When Claude Code presents options, it can feel like you need to pick from a menu. You can tell it to default to free text input instead. Just say "don't give me multiple choice, I'll tell you what I want."

Remap your Caps Lock key. Make it a Control key. This is a small thing that makes terminal life dramatically better. On macOS: System Settings, Keyboard, Keyboard Shortcuts, Modifier Keys. You'll thank yourself.

Understand the "dangerously skip permissions" tradeoff. See Phase 1, Step 1D above for the full explanation of the --dangerously-skip-permissions flag and the clauded alias. Short version: use clauded for day-to-day work, use claude for sensitive operations.

Voice transcription quality can vary. Apple's built-in dictation can regress across OS updates. If you notice accuracy dropping, switch to Superwhisper or Wispr Flow as your primary and keep the other as backup.

Don't try to make it perfect on day one. The MVJ is a scaffold. It will be messy at first. That's fine. The structure will emerge as you use it. Resist the urge to spend three hours designing the perfect folder hierarchy before you've written a single document. Start writing. Reorganize later.

Phase 7: Growing From MVJ to Full Business OS​

The MVJ is the seed. Here's what the growth trajectory looks like.

Step 7A: Week 1​

  • Add 5 more relationship files. Start with the people you interact with most.
  • Write your PRINCIPLES.md: the core decision rules you operate by. What do you value? What are your non-negotiables? What heuristics guide your judgment?
  • Process one real conversation (a meeting, a call, a brainstorm) through the system. See what it's like to capture and route real information.

Step 7B: Month 1​

  • Regular brain dumps are becoming habit. You speak into the system at least a few times a week.
  • Your artifact library is growing: status updates, decision records, relationship files, strategic notes.
  • You're starting to see the compounding effect. Claude Code's briefings are getting noticeably more useful because there's more context to draw from.

Step 7C: Month 3​

  • The system knows enough about you, your operation, and your relationships to generate genuinely useful briefings and catch things you'd miss.
  • You're spending less time trying to remember things and more time making decisions. The recall problem is largely solved.
  • You start to feel the shift: the system is not just a tool you use. It's a thinking partner that operates from your context.

Step 7D: The Organizational Expansion​

When you're ready to bring other people in, the business OS scales with access controls:

  • Each person gets a role-scoped view. Not everyone needs to see everything. The intern doesn't need board-level strategy docs. The sales lead doesn't need HR records.
  • AI agents that act on behalf of team members only have access to documents relevant to their role. This is where the Sovereign Agentic Business OS principles around identity and access management become critical.
  • The vision: a living "company bible" (see Start Your Company Bible) that everyone, human and AI, operates from. Continuously updated. Version-controlled. The single source of truth for how your organization works.

Phase 8: The Meta Work Shift​

Here's the part that feels counterintuitive at first: as your business OS matures, your day starts to look less and less like "work" in the traditional sense.

Low-level execution is increasingly handled by AI. The human's job becomes meta work: defining reality, setting objectives, curating truth, evaluating whether the system is producing good outcomes. Think of it like game design. You are designing the game (the objectives, the rules, the guardrails) and the AI agents are the players executing within those constraints.

A productive day might involve very little typing and a lot of thinking, conversing, and refining the system. You might spend an hour voice-dumping insights from three conversations, review the updated documents, approve the changes, and then ask your business OS to generate a strategic brief. That's not laziness. That is the highest-leverage use of human attention in a world where execution costs are collapsing.

The question is no longer "how do I get all this work done?" It's "am I defining reality accurately enough that the system can do good work on my behalf?"

That's the shift. And it starts with the Minimum Viable Jarvis.

Further Reading​

The best time to start your business OS was a year ago. The second best time is today. Open a terminal. Start talking. The system will grow from there.