How I run my entire business from Claude Code

whats good hustlers,

If you just read one thing about AI this week let it be this. I feel like ive just taken the red pill from morpheus and I can see the future of where all this is headed.

This newsletter will give you the complete lay of the land on where AI is and where you should be focussing your time.

lets get into it!!

📌 TL;DR

GPT-5.4 dropped → It gets its own virtual computer and can click around, navigate apps, and complete tasks for you.
Google workspace CLI → Google shipped a single tool that gives your AI agent access to Gmail, Drive, Calendar, Sheets, and more.
Builder's notes → Built two new skills I'm giving away for free, migrated my entire second brain from Notion to Obsidian, and started using Claude Cowork. Workshop guide coming soon.

Where AI is heading... (perplexity personal computer)

Perplexity Computer is a web based AI agent app that takes a goal, breaks it into tasks, and spins up sub-agents to execute them. You describe what you want, it figures out the steps and does the work.

This is essentially the same thing that Manus, Claude Code, Codex, OpenClaw & Antigravity do...

This week they launched Personal Computer. It runs 24/7 on a Mac mini on your desk with full access to your local files, apps, and emails.

It's basically OpenClaw. Same core idea: a persistent AI agent living on your machine, working around the clock with access to your actual stuff.

But the bigger picture is we're moving from chat to agents, and you don't want to get left behind.

Chat model = question → answer.
Agent = goal → result.

Antigravity, Codex, Perplexity Computer, Manus, Claude Code, OpenClaw. These are all just agent harnesses. Different wrappers around the same concept: take a model, give it tools, memory, and the ability to act on your behalf.

If you understand one of these tools properly, you can use all of them. And I mean that literally.

With all of these harnesses, you're building your assets as markdown files. My Claude Code setup is a folder on my computer that Claude Code runs out of.

Inside my content team folder, I have an instructions file that says "you are my head of content, this is what you do, these are the agents you have access to." Context files about my business, who I am, my brand voice. All my working files.

And then skills, reusable workflows that automate specific tasks. These compound over time and become incredibly valuable.

My /lead-magnet-launcher skill takes a Notion doc and turns it into a full lead magnet funnel. Landing page, n8n email sequence, the lot.

My /weekly-research skill runs 9 days of AI news monitoring across X, Reddit, and the web, and delivers a complete research brief.

Every skill I build works across any harness. I use it in Claude Code. But I can open that same folder in Codex or Antigravity and start working without changing a single thing.

The tools are interchangeable because the assets are the same. Learn one in depth, you've learned them all.

So what makes OpenClaw and Perplexity Personal Computer different?

They add features that give your agent more autonomy. With Claude Code or Codex, when you close the app, the work stops. It's still you giving it prompts, which is great if you're at your desk all day, and most of us are.

OpenClaw blew up because it added three things on top:

Messaging integration (Telegram) so you can work from your phone instead of being chained to your computer
Scheduled tasks that run on a timer without you triggering them (Claude Code and Codex have this too now, but only while your computer's on)
Heartbeat, where the agent wakes itself every 15 minutes, checks for things to do, picks up unfinished work. You can customize that prompt to do specific things on each heartbeat, like check for emails and reply to them

There's a reason everyone wants these running locally. Your files stay on your machine, you control what the agent accesses, and you can wire it into anything you want. Web-based agents are easier to get started with, but you give up a lot of that control.

Perplexity Personal Computer is $200/month (waitlist only right now). It's trying to be the best AI agent harness on the market, taking what OpenClaw gave us but making it enterprise ready and easier to set up. All the AI processing still happens on Perplexity's servers, so you get the always-on local access.

More on where you should start building your own setup below…

Sora Videos API

OpenAI's had the Sora 2 video generation model for a while now, but there was no API. You couldn't integrate it into an app you were building or automate a video workflow. That changed this week.

The new Videos API lets you programmatically generate, edit, and extend videos. Two tiers: Sora 2 at $0.10/second (720p) and Sora 2 Pro at $0.30/second (up to 1080p).

new drop includes:

Character & object consistency. Upload a product, person, or object once, get a persistent character ID, and reuse it across every video you generate. This is the big one.
Longer clips. 16 and 20 second generations (most competitors max at 5-10 seconds).
Video continuation. Extend clips up to 120 seconds by chaining generations together.
Batch rendering. Submit a batch of video jobs and let them process overnight.

In their demo, OpenAI showed a co-branded energy drink campaign: UGC ads in English and Spanish, cinematic product shots, time-lapse videos, all using the same persistent product character across every asset. Check it out here.

Also this week...

Claude now shares context across Excel and PowerPoint → One continuous session across both apps, plus reusable "Skills" for repeatable workflows like variance analyses and slide templates.
OpenAI launched Codex Security → An AI code auditor that finds vulnerabilities, validates them, and proposes fixes.
Karpathy open-sourced AutoResearch → A ~630-line Python tool that lets AI agents run machine learning experiments autonomously while you sleep. In his first overnight run, an AI agent ran 100+ experiments without a single human keystroke. 5,800 GitHub stars in 48 hours.
Gemini overhauled Google Workspace → Docs can now generate full drafts from your existing files and match your writing style. Sheets builds entire spreadsheets from a prompt, including dashboards. Slides generates decks pulling context from your files, emails, and the web. Rolling out in beta now.
Google released Gemini Embedding 2 → Their first multimodal embedding model. You can embed text, audio, images, video, and PDF documents using the same model. It's a little expensive compared to other text embedding options, but video at low fps and audio are really cheap, and the ability to embed all of them at the same time is unmatched. If you've been wanting to build something like "search across all my video content" or "find this moment in my podcast archive" - this is the model that makes it practical.

💡 Builder’s notes

I believe the future of AI is everyone having their own AI OS. A personal operating system built from agents, skills, and automations that compounds every week.

Here's how mine is built out right now.

I sit in Claude Code for all my work. I don't switch between apps. It has access to my Google Calendar, Gmail, Google Drive, Stripe for billing, Notion for project management, Granola for meeting notes (which has been a game changer), plus a bunch more. One central place to do everything.

Let's say I want to send someone an email. I'll just say "hey, send Sabastian an email about XYZ." Claude Code pulls him from my CRM, gets context on him, searches Granola for any previous meetings we've had, and pulls all that in. If the email needs a payment link, it creates one in Stripe and inserts it. If it needs scheduling, it grabs my calendar link, pastes it in, and creates the event. I don't leave Claude Code once.

Same thing with ads. I've got different ad skills I've built to create them. When I'm happy with the creatives, it uploads everything into Google Drive and adds them to my Notion ads database. When it's time to publish, it pulls them straight from those locations. Still haven't left Claude Code.

Then on top of that, I've got skills for all my repeatable processes. A meeting prep skill that runs before every meeting. A morning briefing skill that runs each morning. Skills are just a way to automate your workflows. Instead of building an n8n/ zapier workflow, you build a skill. I have skills across every part of my business that save me 20-40 hours a week.

I made a post on Instagram this week breaking down my top 5 time-saving skills: /lead-magnet-launcher, /youtube-description, /transcribe, /ads-analyst, and /humanise-text. I'll drop the link to download them here.

If you're feeling overwhelmed, here's the checklist. These are the core things to learn, in order. Nail these and you're fully up to speed:

Learn how agents work and pick one harness to master. I'd suggest Claude Code, but Codex or Antigravity will do. Just commit to one and go deep.
Learn about context windows and session management. This is how your agent remembers things across conversations.
Learn AGENTS.md files (called CLAUDE.md in Claude Code). This is the instruction file that tells your agent who it is, what it does, and how it should work.
Learn MCPs to connect your apps. Gmail, Notion, Stripe, Calendar — MCPs are how your agent talks to the tools you already use.
Learn skills to automate your work. Repeatable workflows you can trigger with a single command. This is where the real time savings kick in.
Learn GitHub for sharing and storing code. It's where you'll find other people's skills, configs, and setups — and where you'll back up your own.
Learn sub-agents. This is how you go from one agent doing one thing to a team of agents working together on bigger tasks.

🧰 Tools to try

Proof → A doc editor built for humans and AI agents to collaborate on the same document. Real-time co-editing, comment threads, accept/reject suggestions, and you can see exactly who wrote what (human vs agent). Free, no login. Made by Every.
ComfyUI → Open-source node-based tool for AI image and video generation. Way more control than Midjourney or ChatGPT — you build visual workflows by connecting nodes. Free, runs locally, and NVIDIA just partnered with them for GPU acceleration. If you want to go deeper on AI image/video gen, this is the move.
/btw in Claude Code → Type /btw while Claude is mid-task to ask it a side question without interrupting what it's working on. The answer pops up in an overlay, the main task keeps running. One of those small features that changes how you work.
Firecrawl CLI → Scrape, crawl, and search the web from your terminal. Give it a URL and it pulls clean markdown, HTML, screenshots, whatever you need. Your AI agents can use it too. Great for research workflows and feeding web content into your agent setup.
Twitter CLI → Read your timeline, post tweets, search, bookmark — all from the terminal. No API keys needed. Works with Claude Code out of the box. If you want your AI agent managing your X presence, start here.

i'm not affiliated with any :)

🥣 Brain food

Claude Code + AutoResearch = Self-Improving AI → Nick Saraev takes Karpathy's autoresearch repo (mentioned above) and wires it into Claude Code to build self-improving experiment loops. He demos it optimising cold email campaigns autonomously. Really cool practical application.
AI Agents Full Course 2026 → Another one from Nick. A 2-hour deep dive on building and deploying AI agents using Claude Code, Codex, n8n, and subagent architectures. If you're serious about building your own AI OS, block out time for this.
OpenClaw vs Claude for Running an Agent Team → Brian Casel compares the two ecosystems he uses to run his business. Good companion piece to the main story this week — especially his point about Skills being the key to avoiding platform lock-in.
Standup meetings with your AI engineers → Luke added a scrum meeting to his OpenClaw office. The agents walk into a virtual meeting room and report their progress in real time. This is nuts. 78K views in a day.