Set up your AI operating system with Hyperagent in an afternoon
The system I install in every client's business, stripped down to the fastest version you can build yourself. No code, no dev team, one afternoon.
Watch first19:26In the video, you watched me build an AI operating system (AIOS) into a real business in five days. This page is the version you build yourself, in an afternoon, without writing a line of code.
Quick recap if you skipped it: Arlin runs a retreat business out of Instagram DMs and spreadsheets. We flew into Ibiza, put a proper foundation under it, and stood up a system that reads his data, drafts his outreach, qualifies his applicants and reports to him every morning. I call it an AI operating system or AIOS for short. One system that knows the business instead of a dozen AI tools that don't talk to each other, don't remember anything, and only work when you do.
I install a version of this AIOS system with every client, and I start every single one at the same layer, because the failure mode is always identical: founders bolt AI onto a business it can't see. The tools get blamed, but the tools were never the problem. The foundation was missing. So this guide runs bottom to top, foundation first, and each layer only works because the one under it exists.
Two honest notes before you start.
This kind of system is something I pioneered in February 2026 and built from the ground up on Claude Code. You can absolutely build it there, but it costs a day of API keys and cloud-project setup that ends most beginners before they even get to the complex commands needed to use one properly.
This page uses Hyperagent instead: same shape of system but working minutes after you open it, and it's what we handed Arlin at the end of the week because he doesn't code either. Hyperagent sponsored the video, but the ease of setup it provides is genuinely a game changer for the vast majority of business owners, and I'm extremely happy with my choice to go no-code rather than Claude Code on Arlin's AIOS.
How this page works: every step has a tick box and the page remembers where you're up to, so you can build across a few sessions if the afternoon gets away from you. The tools appear as you need them, with the deals I've arranged where they exist. Nothing on this page is theory. Every prompt is the one I actually ran, word for word.
Context management
No code, nothing to click. You talk, and the system gets a memory.
This first layer doesn't feel like part of the build at all. There's no code and nothing to click, you just talk. What you're building is a context bank: one note that holds everything about your business, so the system knows it without being told again. Every layer that follows stands on this one, which is why I start every client here.
Install Glaido, then open a fresh note on your computer.
Talk your whole business into the note, top to bottom: what it does, your goals, your challenges, your competitors, your ideas, your strengths and weaknesses. It doesn't have to be in any smart order. A stream of consciousness is exactly right.
Export your Claude or ChatGPT data, whichever one you use most. It gets layered into your workspace later in the build, so keep the file next to your note.
Data
The layer most people skip, and the one that separates the top 1% from the top 0.1%.
Right now the most valuable thing your business has is every interaction you've ever had with a customer, and it's locked inside apps your AI can't really use. The second it's in a database you can ask it anything, including the question most owners can never answer: who did I talk to in the past month and never follow up with? Nobody has the manpower to read every conversation by hand. A database plus a workspace does, and that's what the rest of this build stands on.
Do this with whatever your business actually runs on. For Arlin it was Instagram DMs and a spreadsheet of event applications. For you it might be call recordings, emails, your CRM, your inbox. Get it into one place you own and your AI can reach.
Export your data as CSVs. Arlin's two were his event form responses and his Instagram messages, pulled through the Meta export tool that any Meta account can request. Order the export early, it takes a while to arrive.
Go to Supabase and click Start your project. Sign up with GitHub, or an email and password.
Create your organization, then create a project inside it. The free tier gives you two.
You now have an empty project and every instinct says start making tables. Don't. You are not going to touch Supabase again. In layer 04 your workspace builds the database for you and inserts the spreadsheets itself, relationships and all. An account and an empty project is the whole job here.
Integrations
Your apps, reachable without leaving the workspace.
You have context and you have data. Now you need the thing that ties them together. Earlier this year we would do this with Claude Code, and you can absolutely build this there. But as I said before, it costs you a day of hunting API keys, spinning up a cloud project and wiring context management by hand, which is exactly where most beginners stop. Claude Code is a monster PC you build from parts on your desk. It rips, once it boots. This page uses Hyperagent instead: same shape of system, working minutes after you open it.
Sign up to Hyperagent.
Bottom left, click Integrations, scroll to the bottom and find Supabase. Click Connect. Stay in the same browser so you're still logged in, pick the organization you made in layer 02, and authorize it. Your database is now connected.
Connect Gmail the same way. It's one click, and layer 05 needs it to send anything.
Connect whatever else you live in: Slack, Drive, Notion, Stripe, WhatsApp, your call recorder. Each one is a thing you never have to open again.
Workspace
One place that holds the context and reaches the data. This is where the pile becomes a system.
Now go back to that context bank you talked out in layer 01. Create a new project in Hyperagent and load it in, so the workspace starts every thread already knowing your business. Then open a new thread inside that project, because this is where your data arrives.
The relational part is the point, not a detail. Both spreadsheets carry an Instagram handle, so the handle is what stitches an application to the conversation you had with that same person. Get that right and one question can reach someone's answers and their entire message history at once. Name whatever your equivalent is: an email address, a phone number, a customer ID.
What comes back should be one person, their application details, the full conversation thread, and some action points off the back of it. That's the whole foundation proving itself in a single answer. From here the workspace also keeps its own docs updated as you work, so it stays current on what you're doing and what you've decided without you managing it.
Create a new project and load your context bank and your Claude or ChatGPT export into it.
Open a new thread in that project and drag both CSVs straight in.
Switch the model to Opus 4.8. This is a development task, so you want something decent, and stay in plan mode so it plans the build before it touches anything.
Send the prompt below, then read the plan it comes back with and approve it.
Can you please set up a new database in Supabase using our connection that is going to put all of these applications into the database, and then also all of these conversation and DM transcripts, and make sure that they're related to each other based off the Instagram handle in both spreadsheets.
Can you please find a lead in the database and give me a full summary of their application and the conversation I've had with them on Instagram?
Skills
One word kicks off a job that used to take an hour.
Before you automate anything, use the workspace by hand. It sounds backwards, but everything is connected now, so this is where you find out what it can actually do. Ask it real questions. Pull my rep Jimmy's calls from last month and this month, compare them, write me a report. There are no dumb questions here, and the task you end up repeating is the one worth turning into a skill.
Here's the one I built for Arlin, in plain English, no code. I called it Good Morning, or /gm. It reads and analyses every conversation in the database, scores each person on how close they are to buying, and hands back a short list of who's hottest with the first message already drafted for each of them.
What comes back is a small spreadsheet of the current state and ten drafted emails ranked by score. Nothing leaves without you saying so. You can send them all, send all but one, or name a single person. It goes out through the Gmail account you connected back in layer 03.
Start a new session and describe the skill you want in plain English, like the prompt below.
Go back and forth with it until the plan is fully written out, then approve it and let it build.
Save the skill. It appears at the bottom of your project, ready to run.
Open a new session in your project and type /gm. It picks up the skill and runs the whole process.
I want to create a skill that does the following. It reads and analyzes all of the messages in the database and then comes up with the top 10 people based on a hotness score, and also gives me a button so that I can click to send a booking link with a personalized email to each of them.
Arlin took the list, sent the messages the system drafted, and less than 24 hours later a lead he'd forgotten about months earlier turned into a five figure deal.
Liam Ottley
Agents
The work happens without you starting it.
A skill still needs you to press the button. That's being in the loop: same work, faster hands. Most people stop here and think they're done. The last step is handing the trigger over, so the skill runs without you in the room.
I filled out Arlin's form myself. It found my details, scored me 95 out of 100, approved me and sent the booking link by email. That qualifying decision, the one you'd normally make by hand, now happens correctly while you're asleep.
Top of the screen, click Create with chat. Describe the agent you want and let it build on top of your integrations, data and context.
Have it vibe-code the front door too. For Arlin that was an application form he could send anywhere, writing straight into the same database.
Check the system prompt it wrote. Arlin's says: pull in the applicants, research them, save the enrichment, decide if they're a good fit, and email an invite if they're approved.
Open Invocations and choose how it fires. A schedule every 30 minutes works. A webhook is better: the moment an application lands in the database it fires the event and the agent starts. If you don't know how to wire it, describe it in the chat and it walks you through it.
Fill out your own form and watch it run end to end.
A skill will make you faster. But an agent removes the moment you have to actually be there and trigger the skill at all.
Liam Ottley
Every good agent comes out of the same order: explore the workspace by hand, distil the repeated task into a skill, refine that skill while you use it, then push the dialled-in version into an agent. Everyone gets this wrong because they just throw an agent at something, and then wonder why it doesn't work properly.
Apps
A surface you open instead of a chat you prompt.
One more piece. Everything so far still answers you in a chat. The last layer puts a surface on top, so the answer is already there when you look. For Arlin I vibe-coded a leaderboard of his hottest leads straight onto the same database, then put /gm on a nightly schedule so it rescores everyone while he sleeps. As the agent qualifies new applicants through the day, the board updates itself.
Then I handed the whole thing over. Arlin runs it every morning from his phone in the Hyperagent app: he asks who to follow up with, it tells him, he sends the messages. When he wants something new he asks the system for it in plain English. There's no developer on retainer, and he doesn't code.
Ask your workspace for the surface you want. It already has your database, your context and your integrations, so describe the view rather than the plumbing.
Arlin's ranks leads by hotness, has a button to rerun the scoring by hand, and opens in its own tab so he can bookmark it.
Put your skill on a schedule so the data behind the app is fresh before you open it. /gm runs nightly for Arlin, so his board is current every morning.
What you just built
Seven layers, each one standing on the one below it.
Your business, talked out once, so nothing gets re-explained.
Applications and conversations in one place, related on a shared field.
Your data and your apps reachable without leaving one window.
One word reads everything, scores it, and drafts the outreach.
Applications researched, decided and answered without you.
The answer already on screen, rescored nightly.
Founders bolt AI onto a business it can't see, then blame the tools. Every layer here only works because the one under it exists.
The most valuable thing your business has is already sitting in an app you can't search. It isn't more reach.
Explore by hand, distil the repeated task into a skill, then push the dialled-in skill into an agent. In that order.
Arlin runs this from his phone and asks for new things in plain English. No developer on retainer.
Disclosure
Hyperagent sponsored the video this guide comes from. They had no say in the method: this is the same build I run with my clients, and the write-up is mine.
Glaido is an affiliate link. It costs you nothing extra and it's the tool I actually use.
Supabase is here because the build needs a database. No relationship, no deal.