How to build an AI operating system (AI OS)

What is an AI operating system (AI OS)?

An AI OS is the layer underneath your AI tools. It holds your context, your voice, your offers, your customers and your workflows in one structured place that every AI you use can read.

In practice, it's a folder of plain text files: your AI OS workspace folder. It starts on your computer, and a copy in a private GitHub repo gives you a backup, team access and the option to run certain jobs in the cloud. You open that folder in an AI agent, and the agent reads your files at the start of every session, before you type a word.

Here's what that folder looks like:

~/my-ai-os

Claude Codev2.1.138

> /onboard

🧠CLAUDE.mdthe rules, read first every session

·AGENTS.mdsame file, so Codex and Cursor read the same rules

·GEMINI.mdsame file, so Gemini CLI reads them too

·MEMORY.mdwhat came up while working

▸context/who you are, your company, your voice, your brand, your tools

·about-me.md

·my-company.md

·voice-profile.md

·anti-ai-writing-style.md

·design.md

·tools.md

▸decisions/what was decided and why

·log.md

▸inbox/what's open, and your setup answers

·open-work.md

·onboard.md

▸daily-notes/One note per day

▸outputs/everything your AI OS makes, one folder per project

▸projects/briefs and references, one folder per project

▸references/the frameworks, SOPs and tool notes your AI OS runs on

▸assets/logos, headshots, fonts, colors

▸_archive/where deleted things go

·.mcp.jsonthe tools your AI OS can reach

▸.claude/

▸skills/the 12 skills

▸onboard/

▸morning-brief/

▸meeting-notes/

▸research-assistant/

▸slide-deck/

▸visual-explainer/

▸diagram-generator/

▸ai-os-daily/

▸ai-os-connect/

▸ai-os-audit/

▸ai-os-update/

▸ai-os-grow/

▸agents/

·brand-guardian.mdchecks every draft against your voice

That's how I run my own business: an AI OS with Claude Code as the brain.

And it isn't tied to one agent. Each agent reads a rules file when a session starts, and you point all of them at the same one:

Agent	Reads this rules file
Claude Code	CLAUDE.md
Codex	AGENTS.md
Cursor	AGENTS.md and CLAUDE.md
Gemini CLI	GEMINI.md

So you set it up once. Then you can switch agents without starting over.

How an AI agent uses your AI OS

Every AI agent works in a loop. It looks at the situation, decides the next step, takes it, and checks the result. Then it goes around again until the job is done.

Your AI OS is what the agent looks at, every time it goes around. The better that folder knows your business, the better every step.

What an AI OS is not

AI OS vs ChatGPT, custom GPTs and automation tools

Here's how an AI OS compares to the tools you probably use today.

So keep the tools you like. An AI OS sits underneath them, and your agent can run them through connectors.

Why your business needs an AI OS

Without an AI OS, every tool starts from zero every time you open it.

So you paste in who you are. You explain your offer again. You fix the same wrong guesses. And tomorrow you do it all over again.

With an AI OS, the work compounds. Every correction you give, every tool you connect and every skill you add makes the next session better than the last.

Here's what an AI OS does for you, day to day:

And this is still early.

The 7 layers of an AI OS

Every AI OS is built from the same 7 layers. Each layer needs the one below it, so you build them in order.

The test for whether it's working is simple. Open a fresh session and type "write a post about my offer." If it names your offer and your price in your voice, your AI OS is working. If it asks what you sell, start with Context.

Here's each layer in depth.

Layer 1: Context

Context is everything your AI OS knows about you before you type a word: who you are, what you sell, who you sell to, how you write and the rules you work by.

An AI agent with no context gives you generic work, however good the model is. It doesn't know your offer, your customer or how you write, so it guesses. And every guess is something you have to fix.

Most people who give up on AI agents blame the model. In most cases, the agent simply didn't know enough about them.

Think about a new hire. On day 1, they're smart, but they know nothing about your business. You hand them your company docs, and within a week they do the work your way. Context is those docs, for your AI.

Context is a small set of plain text files written in markdown. Markdown is plain text with simple formatting, like headings and lists. It's the easiest format for an agent to read. A PDF or a Word file has to be unpacked first, which is slower and costs more.

It starts with your rules file. That's the one file your agent reads at the start of every session, so it holds 3 things: who you are in a few lines, how you want it to work, and a map of where everything else lives.

Then come your context files. The AI OS Starter Kit uses 6:

Not everything belongs in context. Some things your agent needs every single time. Other things it needs for one task only.

Mix them up and your agent drowns. Load everything into every session, and the one fact that matters gets lost in the noise.

When your AI OS gets something wrong, the cause is almost always one of these 4:

A line earns its place when it changes what the AI writes, decides or refuses.

"I'm passionate about helping creators" changes nothing. "I never use the word journey, and every email ends with a question" changes every draft.

And read every line before you keep it. If your context says your customers struggle with something your product doesn't solve, that line ends up in every page, email and ad your AI writes.

• Everything about the context layer

Layer 2: Data

Data is your business in one place your AI can query: your customers, your sales, your conversations, your content and your meetings.

Context tells your agent who you are. Data tells it what's actually happening in your business: who bought, what they asked, what sold and what worked.

Without data, you're back to copying numbers into a chat. And most business data is locked inside separate apps, where your agent can't see it.

Most businesses have the same 5 kinds of data. Here's what each one lets your AI OS answer:

Sales and customers

Where it might live today

ThriveCartSamCartHighLevelStripePayPal

A question your AI OS can answer

Who bought in the last 30 days, and which offer did they buy?

Email list

Where it might live today

KitFlodeskBeehiiv

A question your AI OS can answer

Which email brought in the most clicks this month?

Courses and community

Where it might live today

SkoolCircle

A question your AI OS can answer

Who joined this week?

Conversations

Where it might live today

Google WorkspaceSlackWhatsAppTelegram

A question your AI OS can answer

What did this customer ask us last time?

Content

Where it might live today

WordPressXLinkedInYouTubeInstagramTikTokFacebookThreadsPinterestSubstack

A question your AI OS can answer

Which pieces brought in the most sign-ups?

Meetings

Where it might live today

GranolaGoogle MeetZoomRiversideOtter.ai

A question your AI OS can answer

What did we agree on the last call?

Data	Where it might live today	A question your AI OS can answer
Sales and customers	ThriveCartSamCartHighLevelStripePayPal	Who bought in the last 30 days, and which offer did they buy?
Email list	KitFlodeskBeehiiv	Which email brought in the most clicks this month?
Courses and community	SkoolCircle	Who joined this week?
Conversations	Google WorkspaceSlackWhatsAppTelegram	What did this customer ask us last time?
Content	WordPressXLinkedInYouTubeInstagramTikTokFacebookThreadsPinterestSubstack	Which pieces brought in the most sign-ups?
Meetings	GranolaGoogle MeetZoomRiversideOtter.ai	What did we agree on the last call?

The data layer does 2 things.

First, it brings your data into one place your agent can query, usually a database. Second, it links the records with one shared ID, such as a customer's email address.

That link is what matters most. It turns 5 separate exports into one answer. Ask about one customer, and your agent pulls their orders, their messages and their history together.

A database like Supabase works well for this. What matters is that it's one place, and your agent can read it without you.

Data is the facts of your business: records, numbers, orders. Memory is what your AI OS learns and wants to recall: decisions, corrections, notes and knowledge. Keep them apart, and each stays easy to search.

Never paste raw exports into your context files. Every line of context loads every session. Raw data belongs in your database, where your agent looks it up only when a task needs it.

• Everything about the data layer

Layer 3: Tools

Tools are the apps your AI OS can reach and use: your email, your calendar, YouTube, your CMS, your payments and more.

Context and data let your agent know things. Tools let it act.

With your tools connected, your agent can check your calendar, read your inbox, draft a reply, pull a report or prepare a post, without you copying anything in.

Here's a simple rule. Every time you paste something into a chat by hand, that's a tool that isn't connected yet.

Here are the kinds of tools most businesses connect:

Website

What you might already use

WordPressFramerWebflowVercelNetlifyCloudflare

What your AI OS does with it

Build, update and publish your pages

Servers

What you might already use

HostingerHetznerDigitalOceanRailway

What your AI OS does with it

Run your automations, agents and apps on a server that's always on

Database

What you might already use

Supabase

What your AI OS does with it

Keep your customers, sales and content in one place your agent can query

Sales and payments

What you might already use

ThriveCartSamCartHighLevelStripePayPal

What your AI OS does with it

Report revenue by offer and look up any customer

Courses and community

What you might already use

SkoolCircleDiscord

What your AI OS does with it

See who joined and what they're asking

Email list

What you might already use

KitFlodeskBeehiiv

What your AI OS does with it

Draft your next email in your voice

Email and calendar

What you might already use

Google Workspace

What your AI OS does with it

Build your morning brief and draft replies

Messages

What you might already use

WhatsAppTelegramiMessageSlack

What your AI OS does with it

Find who's waiting on you and draft the reply

Meetings

What you might already use

GranolaGoogle MeetZoomRiversideOtter.ai

What your AI OS does with it

Turn every call into notes, decisions and next steps

Files and notes

What you might already use

Google WorkspaceNotionObsidian

What your AI OS does with it

Find the document a task needs

Publishing

What you might already use

TypefullyXLinkedInYouTubeInstagramTikTokFacebookThreadsPinterestSubstack

What your AI OS does with it

Draft the next post from what you've published

Research and competitor intelligence

What you might already use

FirecrawlPerplexityApifyScrapeCreatorsRedditFacebook Ad Library

What your AI OS does with it

See what competitors publish, what's working and bring back the sources

Images and video

What you might already use

ChatGPTHiggsfieldMagnificFal.aiOpenArtRunwayHeyGen

What your AI OS does with it

Make images and video from a brief

Video editing

What you might already use

HyperFramesCapCutDescript

What your AI OS does with it

Cut, caption and assemble videos from your script

Audio and music

What you might already use

ElevenLabsSuno

What your AI OS does with it

Make voiceovers and music from a script

What it's for	What you might already use	What your AI OS does with it
Website	WordPressFramerWebflowVercelNetlifyCloudflare	Build, update and publish your pages
Servers	HostingerHetznerDigitalOceanRailway	Run your automations, agents and apps on a server that's always on
Database	Supabase	Keep your customers, sales and content in one place your agent can query
Sales and payments	ThriveCartSamCartHighLevelStripePayPal	Report revenue by offer and look up any customer
Courses and community	SkoolCircleDiscord	See who joined and what they're asking
Email list	KitFlodeskBeehiiv	Draft your next email in your voice
Email and calendar	Google Workspace	Build your morning brief and draft replies
Messages	WhatsAppTelegramiMessageSlack	Find who's waiting on you and draft the reply
Meetings	GranolaGoogle MeetZoomRiversideOtter.ai	Turn every call into notes, decisions and next steps
Files and notes	Google WorkspaceNotionObsidian	Find the document a task needs
Publishing	TypefullyXLinkedInYouTubeInstagramTikTokFacebookThreadsPinterestSubstack	Draft the next post from what you've published
Research and competitor intelligence	FirecrawlPerplexityApifyScrapeCreatorsRedditFacebook Ad Library	See what competitors publish, what's working and bring back the sources
Images and video	ChatGPTHiggsfieldMagnificFal.aiOpenArtRunwayHeyGen	Make images and video from a brief
Video editing	HyperFramesCapCutDescript	Cut, caption and assemble videos from your script
Audio and music	ElevenLabsSuno	Make voiceovers and music from a script

Every tool joins your AI OS through a connector. There are 3 kinds:

In Claude Code, a connected tool's details load only when a task needs them. So adding connectors doesn't weigh down every session.

Start with email and calendar. They power your morning brief, and they let your AI OS learn your voice from emails you actually sent.

Then connect where you publish, then payments, then the rest. Connect one tool at a time, each when a skill needs it.

Connect every tool read-only first. Turn on writing only when a specific skill needs it, and turn it on yourself.

Keep your keys in a local settings file that git ignores. Never in a chat, and never in a context file.

Where money is involved, give your AI its own account and its own key. You can see exactly what it did, and switch off that one key without touching anything else.

• Everything about the tools layer

Layer 4: Memory

Memory is what your AI OS keeps and can recall later. This is where your brain goes.

Every session with an AI agent ends. Without memory, everything you taught it ends with it.

Memory is what makes your AI OS better next month than it is today. Every correction, every decision and everything you learn stays, so you never teach it the same thing twice.

Here's how memory and context differ:

There are 6 kinds of memory in an AI OS:

Your agent may also keep a memory of its own. Claude Code, for example, saves notes per project by itself. Use it, but keep the memory that matters in files you can open and move.

A memory system does 4 jobs:

Your brain lives in its own folder, next to your AI OS, and your rules file points to it by path. Here's the shape:

my-brain/
my-brain/
├── raw/          the sources you drop in: transcripts, articles, notes
└── wiki/         the pages your agent writes and links
    ├── index.md  the map of every page
    └── log.md    what was added, and when

When a body of knowledge keeps growing on its own, like meeting transcripts, give it its own brain. Smaller brains are faster to search and harder to confuse.

Keep one small file of what's active this week, and have your agent read it first. It finds the current thing without searching everything.

• Everything about the memory layer

Layer 5: Skills

Skills are the work your AI OS does.

A skill is an SOP for your AI. It's the same kind of document you'd hand a new assistant: how you do a job, step by step, what good looks like, and what to check before it's done.

You write it once. From then on, one command runs the job your way, every time.

For example, /morning-brief in the AI OS Starter Kit reads your calendar, the email that needs a reply and your open work. It picks the 3 things that matter today and writes it all on one page. You type one command, and you get the finished page.

Without skills, you explain the job every time you ask for it, and you get a different version every time.

With a skill, the job runs the same way every time. And every time you correct it, you fix the skill, so it gets sharper with every run.

Every skill is a folder with a SKILL.md file. The file has 2 parts:

The folder can also hold anything else the skill needs, like reference notes, templates or a script.

Here's how /morning-brief is built:

Your agent doesn't read every skill in full. It starts with just the names and descriptions, and loads the full skill only when your request matches it.

That's why the description matters so much. A vague description means your agent picks the wrong skill, or none at all.

Here's how a skill differs from the other parts of your AI OS:

In Claude Code, custom commands and skills are now the same thing: every skill runs as a slash command. Skills live in .claude/skills/ in your project, or in ~/.claude/skills/ to use them in every project. The AI OS Starter Kit's skills follow the Agent Skills standard.

Here's what makes a good skill:

• Everything about skills

Layer 6: Automations

Automations are what your AI OS runs without you starting it.

You only automate what already works by hand. A skill that fails at 7am with nobody watching is worse than one you run yourself at 9.

So automations sit on top of everything else: context that's been corrected, data that's connected and skills you've already run many times.

Automations run in 4 ways:

Start with one routine before you schedule anything.

Every day, run /morning-brief to start and /ai-os-daily to end. The daily note carries every correction into memory, so tomorrow starts where today stopped.

Every week, run /ai-os-audit to score your AI OS out of 100, then /ai-os-grow. It reads 2 weeks of your notes, finds what you did 3 or more times, and proposes one addition.

When the routine has held for a month, put it on a schedule.

Most business processes don't need an agent deciding things. A skill that runs the same steps at the same time every day is easier to trust, and easier to fix.

• Everything about automations

Layer 7: Apps

Apps show you the answer without you asking for it.

Every other layer answers when you ask. Apps show you what you need before you ask.

I run my own AI OS from a panel like this every day. Everything the other 6 layers produce shows up in one place.

An app can take 3 forms:

Start with the page you already have. /morning-brief writes your brief as one page you open in your browser.

Then build one page for each question you ask every day. Give each page one job, and have a scheduled skill refresh it, so the answer is current when you open it.

You don't have to leave the tools you already use. On HighLevel, for example, you can build a page with Claude and then use it inside HighLevel.

You can also have your AI OS build your own pages on top of the tools you use. On SamCart, for example, a sales dashboard with your own numbers.

• Everything about apps

How to build an AI operating system (AI OS)

Here's how to build your AI OS from scratch, one layer at a time. Each step builds one layer, in the same order as above.

Step 1: Install your AI OS and audit your week

Your AI OS is a folder. You install it once, then open it in the agent you already use.

On your computer, one command does it:

That copies the folder without any git history, so it's yours from the first commit. The script checks that Claude Code is installed, makes the folder a git repo of your own, and opens Claude Code inside it.

Prefer not to use the terminal? Download the ZIP from GitHub and unpack it. Then open the folder the way you normally work:

Every agent reads the same rules. AGENTS.md and GEMINI.md point to CLAUDE.md, so Codex, Cursor and Gemini CLI read exactly what Claude Code reads.

Agents that run in the cloud work differently. Claude Cowork, ChatGPT and Grok Bot can't open the folder the same way, so they connect to your AI OS through its hosted server instead: https://mcp.aios.guide/mcp. Each agent below opens its own site:

Claude CodeClaude CoworkChatGPTCodexCursorGemini CLIAntigravityGrok BotHermesOpenCodeOpenClaw

You need a Mac, Linux or Windows computer, and one of those agents. Claude Code needs a paid Claude plan, Pro or higher, and no API key. The 1-command install needs Node 18 or newer.

Once it's installed, write 2 lists.

The first list is every task you did 3 or more times this week. Those become your first skills and automations.

The second list is where your business actually lives: which app holds your email, your calendar, your customers, your sales, your content and your meetings. That list becomes your plan for data and tools.

Email and calendar

What you might use

Google Workspace

Connected?

Not yet

Sales and customers

What you might use

ThriveCartSamCartHighLevelStripe

Connected?

Not yet

Email list

What you might use

KitFlodeskBeehiiv

Connected?

Not yet

Website

What you might use

WordPressFramerWebflow

Connected?

Not yet

Courses and community

What you might use

SkoolCircle

Connected?

Not yet

Meetings

What you might use

Zoom

Connected?

Not yet

Area	What you might use	Connected?
Email and calendar	Google Workspace	Not yet
Sales and customers	ThriveCartSamCartHighLevelStripe	Not yet
Email list	KitFlodeskBeehiiv	Not yet
Website	WordPressFramerWebflow	Not yet
Courses and community	SkoolCircle	Not yet
Meetings	Zoom	Not yet

Work through these in order, and tick each one off as you go:

Then check that your agent has read the folder:

Prompt 01 · Check it's installed
Read CLAUDE.md and tell me in 3 lines what this folder is and what you will do at the start of every session.

To write both lists faster, let your agent interview you:

Prompt 02 · Audit your week
Interview me one question at a time. First ask what I did 3 or more times this week that I'd hand to someone else. Then ask which tools I use every day, and where my customers, sales, content and meetings live. When you're done, give me both lists as 2 tables.

You're ready for Step 2 when these are true:

Step 2: Build your context

Context is what your AI OS knows about you before you type a word.

It lives in 2 places. Your rules file, and 6 short context files.

Your rules file is CLAUDE.md in Claude Code, AGENTS.md in Codex and Cursor, and GEMINI.md in Gemini CLI. Make them the same file, so every agent reads the same rules.

Use your rules file as a map. It says where everything lives, so your agent goes straight to the right file instead of searching:

CLAUDE.md
## Where things live

- `context/`: who I am, my company, my voice, my brand and my tools. Read every file here at the start of every session.
- `MEMORY.md`: what came up while working. One line per fact.
- `decisions/log.md`: what was decided and why. Append only.
- `daily-notes/`: one note per day.
- `outputs/`: everything you make, one folder per project.

Then the 6 context files. Each one does one job:

The AI OS Starter Kit fills in all of them with one command, /onboard. It takes about 15 minutes.

It connects what it can reach, like your mail and your calendar. It reads your real writing before it asks you anything. Then it asks 12 questions, one at a time, and shows you every file before it saves it.

Answer out loud. Talking gets you more detail than typing, and it sounds more like you. A dictation tool like Wispr Flow or Voicenotes turns it into text.

Don't write a bio. "I'm passionate about helping creators" changes nothing. "I never use the word journey, and every email ends with a question" changes every draft.

When /onboard is done, check what it wrote:

Prompt 01 · Check your context
Read every file in context/ and tell me in 5 lines who I am, what I sell, who I sell to and how I write. Then list anything that is vague or missing.

Then test it on a real task:

Prompt 02 · Prove it works
Write a short post about my main offer, in my voice.

You're ready for Step 3 when these are true:

Step 3: Bring your data into one place

Take your second list from Step 1. For each place your business data lives, export it or connect it.

Keep the tools you use. Connect each one in Step 4, or export its data into your AI OS folder.

Then make sure the records share an ID, like a customer's email address. That link is what lets your agent pull a customer's orders, messages and history together in one answer.

A database of your own, like Supabase, is a later step, for when you want your own HQ.

Prompt 01 · Bring in your data
Here are my exports. Link them by customer email, and show me the full history for one customer, so I can check the links work.

Start with the data you ask about most. For most businesses, that's customers and sales. Add the rest later.

Here's the order to do it in:

Then ask the question most business owners can never answer:

Prompt 02 · Prove it works
Who did I talk to in the past month and never follow up with?

You're ready for Step 4 when these are true:

Step 4: Connect your tools

Start with email and calendar. Those 2 make your morning brief real.

Connect every tool read-only first. Turn on writing only when a specific skill needs it.

Keep your keys out of your files and out of the chat. In Claude Code, the key goes in .claude/settings.local.json, which git ignores, and your .mcp.json points to it by name:

.mcp.json
{
  "mcpServers": {
    "your-tool": {
      "command": "npx",
      "args": ["-y", "your-tool-mcp"],
      "env": { "YOUR_TOOL_API_KEY": "${YOUR_TOOL_API_KEY}" }
    }
  }
}

Each agent keeps its connector settings in its own place:

Agent	Where connectors are set up
Claude Code	claude mcp add, or .mcp.json in your project
Codex	~/.codex/config.toml
Antigravity	~/.gemini/config/mcp_config.json

The AI OS Starter Kit has a command for this, /ai-os-connect. It connects one tool at a time, tells you where the key goes, and runs one read to prove it works.

Where money is involved, give your AI its own account and its own key. You can see exactly what it did, and switch off that one key without touching anything else.

Here's the order to do it in:

To connect a tool, paste this into your agent:

Prompt 01 · Connect a tool
Connect my [tool] read-only. Tell me where the key goes, keep it out of the repo, then run one read to prove it works.

Once your email and calendar are connected, test them:

Prompt 02 · Prove it works
What's on my calendar tomorrow, and which emails need a reply?

You're ready for Step 5 when these are true:

Step 5: Set up your memory system

Memory has a few parts. Set up each one, and your AI OS stops forgetting.

Your corrections go in MEMORY.md, which loads every session:

MEMORY.md
## Corrections

## Preferences

## Infrastructure

Every time you correct your agent, save the correction with the reason. The reason lets it apply the rule to cases you never mentioned. After a correction has held for a month, move it into your rules file.

Your decisions go in decisions/log.md: the date, the decision and why. Your days go in daily-notes/, one note per day.

Your brain goes in its own folder, next to your AI OS, never inside it. Drop a source into raw/, and your agent turns it into pages in wiki/, with an index and a log. Your rules file points to the brain by path.

Claude Code also keeps a memory of its own, per project. Use it, but keep the memory that matters in files you can open.

Prompt 01 · Turn a source into your brain
Read the new file in raw/. Turn it into wiki pages in wiki/, link it to the pages it relates to, add it to wiki/index.md, and log what you did in wiki/log.md.

Keep one source of truth. When 2 notes disagree, like an old price and a new one, archive the old one, or your agent won't know which to trust.

Here's how to set it up:

A week later, check that it remembers:

Prompt 02 · Prove it works
What did I correct you on last week, and what have you done differently since?

You're ready for Step 6 when these are true:

Step 6: Add your skills

Start with the skills that come with the AI OS Starter Kit:

Then build your own. The rule: build a skill the third time you do the same task by hand.

A skill is a folder with a SKILL.md file. The top of the file tells your agent what the skill is for and when to use it:

.claude/skills/weekly-report/SKILL.md
---
name: weekly-report
description: When the user wants the weekly report. Reads this week's daily notes and sales, and writes one page with what happened and what's next.
---

# Weekly report

1. Read this week's daily notes.
2. Pull this week's sales.
3. Write one page: what landed, the numbers, what's next.
4. Check every number against the source before returning.

Keep each SKILL.md under 500 lines, and give each skill one job.

Prompt 01 · Turn a task into a skill
I do this task every week. Here's how I do it: [your steps]. Turn it into a skill with a clear description of when to use it. Then run it once with me, and fix what it gets wrong.

Skills work the same way in Claude Code and Codex. Both follow the open Agent Skills standard: a folder with a SKILL.md file.

Claude Code reads skills from .claude/skills/ and runs them with /skill-name. Codex reads them from .agents/skills/ and runs them with $skill-name.

To install a skill someone else wrote, copy its folder into .claude/skills/ in Claude Code. In Codex, use $skill-installer.

In Claude.ai and the Claude desktop chat, you upload a skill as a zip. Turn on Code execution and file creation in Settings, Capabilities. Then go to Customize, Skills, click the plus button, and choose Upload a skill.

A plugin is several capabilities in one install. It can bundle skills, sub-agents and hooks, and it can bring its own tools too.

In Claude Code, add a plugin marketplace, then install from it. The AI OS Starter Kit is 1:

Claude Code
/plugin marketplace add thenavidm/ai-os
/plugin install ai-os@ai-os

Anthropic's own catalog, claude-plugins-official, is already there. Browse it with /plugin, or at claude.com/plugins.

In Codex, open the Plugins tab, or type /plugins in the CLI, and select the plus button to install.

Only install plugins you trust. A plugin can run code on your computer.

Other agents handle skills and plugins their own way. Check each agent's own docs before you copy a folder in.

Here's how to build your first skill:

Then put your new skill to work:

Prompt 02 · Prove it works
Run my new skill on this week's work. Then tell me what you would change in it.

You're ready for Step 7 when these are true:

Step 7: Automate what already works

Only automate a skill after it has worked by hand for a few weeks.

Start with a daily and a weekly routine. Every day, run /morning-brief to start and /ai-os-daily to end. Every week, run /ai-os-audit to score your AI OS out of 100, then /ai-os-grow to add one thing.

Then put what works on a schedule. In Claude Code, you have 3 ways:

A cloud routine starts from a fresh copy of your repo, so it can't read files that only exist on your laptop. Put what it needs in the repo, or run it as a desktop task.

Here's how to get there:

When a skill is ready, put it on a schedule:

Prompt 01 · Schedule it
Schedule /morning-brief to run every weekday at 7am as a desktop task, so it can read my files. Then show me where to check that it ran.

After a week, check what ran without you:

Prompt 02 · Prove it works
What ran on its own this week, and what did it produce?

You're ready for Step 8 when these are true:

Step 8: Add your apps

The first app is already there. /morning-brief writes your brief as one page you open in your browser.

From there, build one page for each question you ask every day: your numbers, your pipeline, your content. Each page refreshes on its own, so the answer is waiting before you ask.

Prompt 01 · Build your first page
Build me one page that shows this week's sales, my open work and my next 3 calendar events. Refresh it every morning at 7, and tell me how to open it on my phone.

Then make that page part of your day:

The next morning, check it:

Prompt 02 · Prove it works
Open my page and tell me what changed since yesterday.

Your AI OS is built when these are true:

Your AI OS action plan

Here's your AI OS action plan: every step of building your AI operating system (AI OS) in one list, so you can tick it off as you go. It follows the 8 steps above.

1. Install and audit

2. Context

3. Data

4. Tools

5. Memory

6. Skills

7. Automations

8. Apps

Your first 14 days

Here's what to expect in your first 2 weeks, and the month after.

Most first audit scores are low, and that's fine. The number climbing is the point.

AI OS KPIs: how to measure your AI OS

These are your AI OS KPIs: one number for each layer that tells you whether it's working.

You can't improve what you don't measure. Here's one thing to track for each layer:

On top of that, /ai-os-audit scores your whole AI OS out of 100 every week, and puts it in one of 4 stages:

Common AI OS mistakes

These are the mistakes that hold most AI OS setups back, and the fix for each 1.

Writing a bio instead of a brief

"I'm passionate about helping creators" changes nothing your AI writes. Neither does your life story.

What changes the output is facts it can act on: your offers and your prices, who buys them, the words you never use, how every email ends.

The fix: go through each context file and delete any line a competitor could paste onto their own site. Replace it with a name, a number or a rule.

Putting everything into context

Context loads in every session. So every line you put there is read every time, whether the task needs it or not.

Put too much in, and the right piece gets lost in the noise.

The fix: keep context to what every session needs: who you are, what you sell, your voice and your rules. Everything else goes into memory, where your agent looks it up only when a task needs it.

Pasting the same data in every day

Every time you paste in the same numbers, notes or customer details, you're doing the job a connection should do.

The paste is also where mistakes creep in. You grab last week's file, or half of a list.

The fix: treat every repeat paste as a missing connection. Connect that tool in Step 4, read-only first.

Giving your AI your personal account

With your own login, your AI can do anything you can do. And when something goes wrong, you can't tell what it did from what you did.

The fix: give it its own account and its own key, read-only first. You see exactly what it did, and you can switch off that one key without touching anything else.

Automating before it works by hand

A schedule runs a skill whether its output is right or not. Automate a skill that's still wrong, and you get the wrong result every morning, on time.

The fix: run it by hand for a few weeks. When you stop fixing its output, put it on a schedule in Step 7.

Treating it as finished after setup

Setup is day 1. Memory pays off after weeks, as your corrections pile up and your agent stops repeating the same mistakes.

The fix: run /ai-os-daily at the end of every day. Once a week, run /ai-os-audit to see your score, then /ai-os-grow to add one thing.

Your AI OS summary

That's the full system, one layer at a time.

If you followed the steps, here's what you built:

• Your workspace: one folder, installed and open in your agent
• Your context: 6 files that tell every session who you are and how you write
• Your data: the records you ask about, connected or in one place
• Your tools: your email, your calendar and the apps you live in, connected read-only
• Your memory: corrections, decisions and daily notes that don't get lost
• Your skills: the starter kit's skills, plus one of your own
• Your automations: your morning brief, waiting before you sit down
• Your apps: one page that answers before you ask

Each layer only works because the one under it exists.

What matters most

Context comes first. Every layer after it works better because your AI already knows you.

Keep the tools you use. You don't switch anything to build an AI OS. You connect what you already have.

Build it by hand, then automate it. A skill earns its schedule by working first.

Final thoughts

Your first setup won't be perfect. Mine isn't either.

Update your files every time you catch something that's off. The more you run it, the more it knows, and that's the part that compounds.

Where to go next

Each layer has its own page, with more depth than this guide:

• Context: what goes into your context, and how to keep it lean
• Data: how to bring your business data into one place
• Tools: how to connect your apps safely
• Memory: how to build a memory system that never forgets
• Skills: the skills that come with your AI OS, and how to build your own
• Automations: how to run your AI OS without you
• Apps: how to see the answer without asking
• MCP & CLI: how to connect your AI OS to any AI agent

Start with the layer where you lose the most time.