Build your own Jarvis.
An AI operator that lives on your machine, knows your voice, reaches your email and calendar, and quietly runs the daily busywork. Built on tools you already pay for — live in 30–60 minutes.
Most people use AI like a vending machine. Type a question, take an answer, close the tab. That was the right idea in 2023. Today it leaves nearly all of the value on the table.
The shift is to stop visiting the AI and start giving it a home — a place to live, a set of tools, and a memory. Once it has those three things, it stops being something you check and becomes something that works for you in the background. That is the whole idea behind a Jarvis, and you do not have to build any of it from scratch.
text it
knows your voice + history · uses your email + calendar
handed back
drafts your replies · turns voice notes into done · 7am brief · never drops a follow-up
What you'll have when it's running
Replies, drafted and waiting. New mail gets sorted and answered in your voice — you approve or edit, then send. No blank page.
Voice notes turned into done work. Say what you need into your phone between jobs; it comes back as the finished email, post, or quote.
A brief every morning. Your day, your open loops, and the one or two things that actually need you — before you've had coffee.
Nothing dropped. Every promise you make gets tracked and surfaced before it slips. That one alone usually pays for the whole thing.
Built on tools you already pay for. Roughly 30–60 minutes to get it live; about a week to dial in.
The one decision that matters.
Run it on Claude Code, on the Max plan. That is the decision. Everything after this is detail.
Claude Code is not a chat window — it is a working environment. It has a project folder it can read and write, commands it can run, and outside tools it can reach. That is what lets an assistant cross the line from "answers questions" to "gets things done while you're on a job." Setup is light: it installs in a couple of minutes and runs on an ordinary Mac, Windows, or Linux machine. There is no GPU and no heavy hardware involved — all the model work happens on Anthropic's servers, so your computer is only sending and receiving.
The Max plan is what makes it sane to leave running all day. It is a flat monthly fee with a generous usage budget, so you stop counting tokens and just let the thing work. One honest note, because you will hear "unlimited" thrown around: it is a large fixed budget, not literally infinite — but for a personal assistant that runs your day, you will not come close to the edge.
Install and sign in.
Use the native installer from the official download page (claude.com/download) — it needs no extra runtimes and keeps itself updated. If you prefer the package manager and already run Node 18 or newer, the alternative is:
npm install -g @anthropic-ai/claude-code
Then, from any folder, run claude. The first run opens your browser and asks you to sign in to your Anthropic account. Sign in with the account that carries your Max plan and you are authenticated — no API keys to manage.
Three honest answers before you start.
“Isn't this just ChatGPT?” No. A chatbot answers and forgets. This one has a home, your files, and a memory — it works in the background and remembers what it did yesterday. Same models underneath; a completely different relationship.
“Do I need to code?” No. Everything you write is plain English in text files. You'll run two or three commands by copy-paste — that's the whole of it.
“Is my data safe?” Your files stay on your machine. The model only sees what a task needs, and the rule it never breaks is that it drafts — it never sends or shares anything without you. For truly sensitive work, the whole setup can run on a machine you own (Tab 5).
Everything your Jarvis knows and does lives inside one project folder. Get this structure right and the rest of the build is almost mechanical — you are mostly filling in files.
your-jarvis/ ← the project root (keep it under version control)
├── CLAUDE.md ← who it is, your voice, the map, the rules
├── knowledge-base/ ← what it knows about you and your business
│ ├── about-you.md ← who you are, what you do, how you work
│ ├── voice.md ← how you write, with 5–10 real samples
│ └── faqs.md ← your prices, policies, the answers you repeat
├── pipeline/ ← open loops: who you owe what, and when
│ └── follow-ups.md
├── content/ ← posts, newsletter, anything you publish
│ └── drafts/
├── templates/ ← the messages and documents you write over and over
├── comms/ ← working drafts + the daily follow-up queue
│ └── drafts/
├── logs/ ← a record of what it did
│ └── daily/ ← morning briefs and evening wraps land here
└── .claude/
└── skills/ ← the small instruction files that make it actYou can create the whole thing in one command:
mkdir -p your-jarvis/{knowledge-base,pipeline,content/drafts,templates,comms/drafts,logs/daily,.claude/skills}
cd your-jarvis && git initThe CLAUDE.md is the heart of it.
CLAUDE.md is a plain document at the root that tells the assistant who it is, how you write, where everything lives, and the rules it must never break. Claude Code reads it automatically every time it starts. This is the file you will keep editing for the first week — and every edit makes it more like you. Here is a starter you can paste and fill in:
# CLAUDE.md — [Your Name]'s Jarvis ## Who you are You are my personal operator. You handle the busywork so I don't. You are calm, brief, and dry. You never use hype, exclamation points, or filler. When you are unsure, you ask one short question rather than guessing. ## How I write My writing voice is in `knowledge-base/voice.md`. Match it exactly. Default to short. No corporate tone. Read the samples before drafting anything. ## What lives where - Who I am and what I do: `knowledge-base/about-you.md` - My prices, policies, repeated answers: `knowledge-base/faqs.md` - Open loops I owe people: `pipeline/follow-ups.md` - Templates for things I write often: `templates/` - Where you stage drafts: `comms/drafts/` - Where you log your work: `logs/daily/` ## The rules (never break these) 1. NEVER send an email, message, or reply on your own. Draft it, show me, wait for "send." 2. Only touch the specific file or tool a task needs. Do not roam. 3. Log what you did, with a timestamp, in today's file under `logs/daily/`. 4. If a request is unclear, ask one question. Do not assume. 5. Never share my private information with anyone but me. ## What you do on your own - Each morning at 7am: write today's brief (see the morning-brief skill). - When I drop a voice note or message: turn it into the finished thing I asked for. - When new email arrives: sort it and stage drafts — but do not send. - Always: track open loops in `pipeline/follow-ups.md` and remind me before they slip.
voice.md — the file that does the most work.
If you perfect one file, make it this one. voice.md is where it learns to write like you instead of like a press release. The trick isn't describing your voice — it's showing it. Paste 5–10 real samples: actual emails, texts, captions you've already sent. Then a couple of lines on the patterns.
# voice.md — how I write
## The rules of my voice
- Short. Most replies are two to four sentences.
- No filler openers ("Hope you're well!"). Get to it.
- Dry, a little blunt. Periods, not exclamation points.
- Lowercase is fine in casual replies. Never corporate.
## Real samples — paste 5–10 of your own
[a client email you actually sent]
Hey Mark — got the files. Tuesday works.
I'll send a quote by end of day. Anything change on scope?
[a quick text]
yeah that's fine. send it whenever.
[a post caption]
Most people automate the wrong thing first.
Start with the task you dread, not the one that's easy.
## What to avoid
- Don't make me sound excited. I'm not.
- Don't explain things I didn't ask about.
- Never "delighted," "reach out," "circle back," or "synergy."That's the brain. What turns it into an operator that actually drafts your replies, works your voice notes, and briefs you every morning is the build — the steps, the tool connections, and the first skills. That part is members-only, below.
members only · the build
The build, step by step.
The full ten-step build, the two-tool rule, and your first three skills — copy-paste ready. Members of the Austin AI Club unlock the build, the first wins, and the rest of the library.
Already a member? Email usThis is the whole thing, in order. Do not skip step nine.
- 01
Install Claude Code and sign in with your Max account.
From Tab 1. No keys.
- 02
Create the project folder and run git init.
From Tab 2 — so the assistant's brain has a history you can roll back.
- 03
Scaffold the folders.
Use the one-line
mkdircommand from Tab 2. - 04
Write CLAUDE.md from the starter.
Spend real time on the rules section.
- 05
Fill three knowledge files.
about-you.md— who you are, what your business does, how you work.
voice.md— paste 5–10 samples of your actual writing (real emails, texts, posts). This single step does more for quality than anything else.
faqs.md— your prices, policies, and the answers you find yourself repeating. - 06
Connect two tools only — your email and your calendar.
Use Claude Code's tool/MCP connection for these. Resist adding more; every extra connection adds weight and noise. (Check the current connection command in the Claude Code docs.)
- 07
Add your first skills.
Small instruction files in
.claude/skills/(examples below). - 08
Test it by hand.
Ask it to draft one reply. Ask it to summarize your day from the calendar. Watch where it sounds wrong, and fix
voice.md. - 09
Live in it for a week — personal layer only.
Use it daily for real work before you trust it with anything important. Tune
voice.mdandCLAUDE.mduntil the output is indistinguishable from you. This is the step people skip. Don't. - 10
Then, and only then, go deeper.
Tab 5.
Your first three skills.
A skill is just a short markdown file telling the assistant how to do one repeatable job. Drop these in .claude/skills/ (one folder each — confirm the exact layout in the Claude Code docs, the format evolves). Here is the first one in full:
# draft-reply Use this when I ask you to reply to an email or message. Steps: 1. Read the message I'm replying to in full. 2. Read `knowledge-base/voice.md` and match my voice exactly. 3. Check `knowledge-base/faqs.md` for any facts, prices, or policies that apply. 4. Write the reply — short, in my voice, no filler. 5. Save it to `comms/drafts/` and show it to me. Do NOT send it. 6. Log what you drafted in today's file under `logs/daily/`.
The other two follow the same shape:
- 02voice-note-to-output — a spoken or messy note becomes the finished email, post, task, or quote, in your voice, staged for approval
- 03morning-brief — reads your calendar + follow-ups and writes a short brief to logs/daily/: what's on today, what you owe, the one or two things that need you
Rather not build it yourself? Reach out directly at austinaiguy.com
members only · the payoff
The first wins.
The four things working in week one — inbox triage, voice-note to done, the 7am brief, follow-up memory — and exactly how to trigger and tune each. Join to unlock.
Already a member? Email usThis is what you're actually here for: the basic things, working, in the first week. Get these four right and the rest takes care of itself.
Inbox triage and drafts in your voice.
The assistant reads new mail, sorts it by what actually needs you, and stages a reply for anything routine — written in your voice, waiting for one tap to send. You are not writing from a blank page anymore; you are approving or editing. The rule from your CLAUDE.md holds: it drafts, you send. Nothing leaves without you.
How to tune it: if a draft sounds off, the fix is almost always in voice.md. Add the email it got wrong, then add a version rewritten the way you'd actually say it. A few of those and it locks onto your voice.
Voice-note → finished output.
Between jobs, in the truck, walking to the car — you say what you need into your phone and send it the note. It figures out what you actually want (a reply, a quote, a post, a reminder), produces the finished version, and stages it for you. The messy thought you'd have lost becomes done work.
The 7am brief.
Every morning it reads your calendar and your open loops and writes you a short brief: what's on today, what you owe people, and the one or two things that genuinely need your attention. No dashboard, no app — it just shows up. This is the habit that makes the whole thing feel alive.
Follow-up memory.
Every promise you make — "I'll send that quote," "circle back Thursday" — gets logged to pipeline/follow-ups.md. Before anything is about to slip, it tells you. This one quietly saves more money than the rest combined, because the leads and clients you lose are mostly the ones you simply forgot to chase.
live, drafting by hand
voice locked in
briefing you every morning, follow-ups tracked
Rather not build it yourself? Reach out directly at austinaiguy.com
members only · no ceiling
Going deeper.
How the same folder base grows into business comms, a content engine, a real pipeline, and an on-premise setup for anything sensitive. Join to unlock the rest.
Already a member? Email usOnce the personal layer earns your trust — and it will, fast — the same folder base grows into the parts that run a business. The build is identical; you are only adding knowledge and skills.
Business comms. Outreach drafted per contact, follow-up sequences that run themselves, replies staged for your approval.
A content engine. Posts, your newsletter, and guides drafted on a schedule and queued for one-tap approval.
A simple pipeline. Every lead and client tracked by stage, with the assistant flagging who's gone quiet before you'd notice.
On-premise, for anything sensitive. If you handle records that can't touch the cloud — medical, legal, financial — the entire setup can move onto a dedicated machine in your own office, under your own lock.
This is the point where most people realize there isn't a ceiling — only the next thing you decide to hand it.
Rather have it built and kept sharp for you? Reach out directly at austinaiguy.com.