Bonjour mes amis,

My weeks somehow get crazier and crazier, and we’ve really outdone ourselves this week.

I recorded the best FREE course on AI agents with Greg Isenberg + lots more, lets get into it shall we..

📌 TL;DR

• I did a pod with Greg Isenberg → Sat down with the biggest startup podcast to give everyone a full course on AI agents. Go watch it.

• Anthropic shipped Dispatch and Channels → Two new ways to talk to Claude remotely from your phone. Channels is the Claude Code equivalent via Telegram or Discord.

• Builder's Notes → My honest take on where things are at. Plus a breakdown of my /youtube-publish-workflow skill that creates youtube thumbnails.

• My top 6 MCPs → The exact MCP setup I use every day.

• Also this week → GPT-5.4 mini and nano, Mistral Small 4, Midjourney V8, MiniMax M2.7, and Lovable trying to be everything to everyone.

I did a pod with Greg Isenberg

Bit of a surreal one this week.

I did an episode with Greg Isenberg on The Startup Ideas Podcast - probably the biggest podcast in the AI & startup space right now.

I've been watching Greg for about 12 months. All of a sudden, there I was sitting across from him. Wild. He's an amazing bloke and exactly how he comes across in the videos.

I wanted to make this a full course on agents. Where things are at, how they work, and how you can start building your own. Because most founders are still just messaging chat models; asking questions, getting answers. The ones who start using agents to actually delegate real work are going to pull so far ahead that catching up becomes really hard.

No matter what level you're at, there's something in here for you. Give it a watch.

This was a massive milestone for me. Really appreciate all the positive feedback in the comments. Means a lot.

Delegate tasks to Claude Cowork from your phone

Anthropic had a massive week. 1M token context is now just the default across Opus and Sonnet (Claude’s main AI models). Your context window is basically how much your AI can hold in its head in one chat without it losing the plot.

Anthropic also launched Dispatch. Dispatch creates one persistent conversation between the Claude mobile app on your phone and the Claude Desktop app on your computer.

It's not Claude Cowork on your phone. Your phone is the messaging interface. Your computer is the engine. Everything Claude does, every file it reads, every connector it accesses, every skill it loads, happens on your desktop. Your phone just sends the instructions and receives the results.

Think of it as a walkie-talkie to a computer that's already running. A remote control if you will.

It covers 90% of what I was using OpenClaw for, but feels far less likely to upload my entire drive to a malware site.

Felix Rieseberg also announced that Dispatch can now launch Claude Code sessions, using the same underlying tech as its remote control feature. So the direction is clear. AI agents running on your computer, doing real work, while you delegate from wherever you are.

Now, Anthropic are known for shipping features early to gather feedback fast. Dispatch certainly has some rough edges and some things don't work amazingly, but the core is solid. It's early software, so set your expectations.

But "early" in the Claude ecosystem tends to mean "transformative within 60 days." Cowork itself went from rough preview to daily workhorse in under eight weeks.

Things to note:

• Your existing Cowork skills, context files, and connectors all carry over automatically

• Pre-configure Gmail, Slack, Notion connectors on desktop first, Dispatch inherits them

• Dispatch doesn't read your CLAUDE.md unless you tell it to. Either include "read CLAUDE.md" in your message, or add a line in your global instructions to always find and read CLAUDE.md when sent a task via Dispatch. I assume this will be fixed very soon

• No notifications when tasks finish, you have to check manually

• One thread only, no way to manage multiple conversations

• "Keep awake" must be toggled on in your Claude Desktop app or Dispatch dies when your Mac sleeps

• No file picker on mobile, describe the folder path

• Can't attach files from your phone yet

I like this one though. Will definitely be using it.

Connect Claude Code to Telegram, Discord and more…

About an hour ago, Anthropic also shipped Channels for Claude Code. Telegram and Discord are included.

You set up a bot, pair it with your Claude Code session, and you can message Claude from your phone while it works on your computer.

The catch is your computer needs to be on and Claude Code needs to be running. If the session is closed, messages don't queue up. They're just lost.

For always-on use, you leave Claude Code running in a background terminal. And if Claude needs permission to do something (like editing a file), the session pauses until you approve it at your computer.

When someone asked Thariq (from Anthropic) how Channels is different from Dispatch, he said: "We just want to give you a lot of different options in how you talk to Claude remotely." Dispatch is the default for most people for Claude Cowork. If you're a Claude Code user and want to message it from your phone, Channels is probably the way to go.

Things to note:

• Research preview, things may change

• Requires Claude Code v2.1.80+ and a claude.ai login

• You can send files back and forth (images, documents, up to 50MB on Telegram)

• Only people you've explicitly approved can message your bot, everyone else is silently ignored

After testing… Dispatch is solid out of the box for everyday users to chat to their Agents from their phone, and Channels is a bit more difficult to setup and still very new and janky.

Like I said, Claude are building OpenClaw one feature at a time. But in this transitionary period while they are shipping these features, they are still very rough around the edges.

Also this week...

• OpenAI released GPT-5.4 mini and nano → smaller, faster, cheaper versions of GPT-5.4 built for high-volume agent workflows. Mini approaches the full model on most benchmarks at a fraction of the cost.

• Mistral launched Mistral Small 4 → a 119B open-weight model that unifies their reasoning, vision, and coding models into one. Apache 2.0 licensed, free for commercial use.

• Midjourney V8 Alpha is live → 5x faster than V7 with native 2K resolution. No more upscaling. Character references and image prompts aren't supported yet, but the speed alone is a big upgrade.

• MiniMax launched M2.7 → The benchmarks are strong too, scoring near Claude Opus and GPT-5.4 on agentic coding tasks at roughly 5% of the cost. One thing to note: unlike previous MiniMax models which were open-weight, M2.7 is proprietary and API-only. It's only two days old so nobody's really stress-tested it yet, but worth keeping an eye on.

• Lovable announced they're expanding beyond app building into data science, business analysis, deck building, and marketing. Their words: "a general-purpose co-founder that can do anything." Personally, I think this is a mistake. Trying to be everything to everyone usually means you end up being nothing to anyone. Lovable was great because it was laser-focused on one thing. Why would anybody use this over Claude Code or Gemini for general tasks? I'm not sure.

💡 Builder’s notes

First things first. If you haven't watched my episode with Greg Isenberg yet, go watch it. It's literally everything you need to understand AI agents.

Now, my honest take on where sh*t's at.

It feels like everything is half-built (because it kind of is). OpenClaw, Perplexity Computer, Manus. None of them feel like “finished” products yet.

But if I had to pick one to bet on right now, it's Claude Code. This is where all my agents live. Telegram integrations, scheduled tasks, agent teams.

The new messaging features like Channels and scheduled tasks are a bit technical to set up, but no more technical than OpenClaw was. And they're only going to get better.

I think the whole AI world is a bit wobbly right now. But you know what's a 100% safe bet, no matter what?

• Building your skills. Reusable workflows that automate specific tasks.

• Building your AGENTS.md files. The instruction files that tell your agents who they are and how to work.

• Building your context files. Your ideal customer profile, about me, offers, brand voice.

These are the assets that travel with you. If next week Perplexity Personal Computer drops and it's incredible, it'll take me about an hour to move my entire system there. Or to Manus. Or to wherever. Because the assets are the same. The tools are interchangeable because everything you're building is just markdown files in a folder.

So if you're feeling paralysed about which tool to pick, stop. Pick one, I suggest Claude Code (or cowork if you’re really not tech savvy), and start building your assets, and know that nothing you build is wasted :)

AND BOY HAVE I BUILT SOME INSANE AGENTS AND SKILLS!

I built a skill called /youtube-publish-workflow that takes me from finished video to live on YouTube without leaving Claude Code.

It's an orchestrator skill, meaning it calls 5 of my other skills in sequence:

1. /transcribe → Transcribes the full video first so everything downstream is working from the actual content.

2. /youtube-titles → Uses the transcript + competitor research to generate 20 title options. I pick 3, then narrow to 1.

3. /youtube-thumbnail → Creates 4 thumbnail concepts using Gemini with competitor thumbnails as style references. I pick a direction and iterate until I'm happy.

4. /youtube-description → Generates timestamps, researches links for every tool I mention, and writes the full description.

5. /youtube-upload → Uploads everything to YouTube via the Zernio API.

Here are some of the thumbnails it's been spitting out (my face is a bit cooked but its a work in progress 😭):

My recent YouTube video was published end-to-end with this workflow. Title, thumbnail, description, timestamps, upload. One command.

The thumbnail skill has its own brand guidelines file and a folder where you upload reference images of thumbnails you like. It studies your style and what's already working in your niche, then produces concepts that match. You set it up once and every thumbnail after that is on-brand.

I'm going to be making this skill available for you guys over the next week or two.

The whole workflow takes 15 minutes depending on video length. Publishing and thumbnails used to be a 2-3 hour process. Now I say "new YouTube video ready" and Claude handles it.

Skills compounding. Every one you build makes the next week faster than the last.

🧰 Tools to try

I just dropped a video breaking down the six MCPs I use every single day. If you're using Claude (or any AI agent) without MCPs, you're leaving so much on the table.

1. n8n MCP → Connects Claude to n8n (workflow automation, like Zapier). Claude plans the workflow, builds it, and can even trigger it to run. I barely open n8n myself anymore.

2. Nano Banana MCP → AI image generation inside Claude. I use this for e-com product shoots and ad creatives (and the youtube thumbnail workflow).

3. Firecrawl MCP → Web scraping. Claude can search the web by default, but it can't actually extract things from a website. Firecrawl fixes that. Landing page audits, pulling branding from competitor sites, scraping Google Docs. Anything where you need the actual content of a page.

4. Perplexity MCP → Research without leaving Claude. No more jumping out to a separate Perplexity tab, losing all your context, doing the research, and bringing it back. Claude just fires off Perplexity searches mid-conversation. This is how my weekly newsletter research gets done.

5. Apify MCP → A library of thousands of online scrapers. X, TikTok, Google Maps, Reddit, e-commerce, lead gen. If you could only set up one MCP, this would be the one.

6. Playwriter MCP → Gives Claude access to your actual browser. Huge for anything that doesn't have an API. Works alongside Firecrawl too. Between the two of them, there's not much on the web you can't reach. Because the Claude browser extension kinda sucks for taking actions on your browser.

🥣 Brain food

• Anthropic is building a Claude pixel avatar creator → Jane Wong (the woman who scoops every tech company) spotted this in the Claude app. It's fun to see Anthropic building personality features alongside the heavy agent stuff.

• How to Use Claude Dispatch to Run Your AI While You Sleep → If you read the Dispatch section above and want a step-by-step setup guide, Nav Toor wrote a great one. Worth reading.

• My Unlikely Friendship with the OG of Film Photography → wholesome asf

• I gave my friend an intervention (he's not an addict) → this was funny