Hi all,
Just landed back in Copenhagen after a lovely few days visiting my parents in London.
I have a BIG month ahead and couldn’t think of a place I’d rather build from, than this apartment and desk…
I just bought a monitor today to complete the setup!
I’m also trying something new this newsletter; backlinking all AI terms to a glossary page with my breakdown of what they mean. (things like; Context windows, MCP’s, Sub-agents, CLI’s, etc)
I want this newsletter to cater to people of all AI literacy levels…
📌 TL;DR
Muse Code → Meta entered coding agents with a dirt-cheap model. The catch is why I’m sticking with Anthropic and OpenAI.
Wispr Flow Notetaker → Wispr added meeting notes to the voice-to-text app I already use. One test exposed a serious mistake.
Agent Plugins → Experts can now package their knowledge into something anyone can install in an agent. Claude made it difficult.
Cloudflare OS → An open-source workspace for shared skills, apps, workflows and any model. I think every company ends up with one.
Builder’s notes → I built a Kindle library, a newsletter bookshelf and finally found a way to use Opus 5 without talking to it.
Granola is shaking…
FUCK YES BABY.
Wispr Flow just launched Notetaker, and I love this.
I've been waiting for Granola to add voice-to-text. Instead, Wispr Flow has added meeting notes. I'm desperate to consolidate the two tools.
There's no bot joining the call. It listens through your Mac, labels speakers, and gives you a summary and action items. You can also connect it to Claude or ChatGPT through the official MCP they released.
I ran Wispr and Granola on the same meeting.
Granola's notes were slightly more accurate in that one test: Wispr said I'd recommended Gooseworks to a prospect when he'd actually recommended it to me, which is a pretty big mistake.
But one meeting is nowhere near enough to declare a winner. I imagine they're probably similar overall.
Wispr Notetaker; action items broken down per person
I still like how Wispr separates everyone by name and pulls out action items. I'll most likely consolidate into Wispr anyway, unless Granola releases voice-to-text ASAP lol.
Notetaker is Mac-only for now
Prepare for a plugin invasion.
When Custom GPTs launched, every corner of the internet made one.
Coaches built client assistants. Agencies built copy bots. Recruiters built CV screeners. Estate agents built property matchers. Course creators handed them out as lead magnets.
Custom GPT’s
Expect the same thing with plugins…
At the simplest level, a plugin is a bundle of skills and/or MCPs that you can add to an agent. Skills teach the agent how to do repeatable jobs. MCPs connect it to other tools and data.
An accountant could make one that sorts receipts and updates a spreadsheet.
A marketer could package their process for researching competitors and writing ads.
A consultant could let clients use their entire method inside their own AI.
Before this, sharing that stuff usually meant sending someone a zip file or telling them to use GitHub. Normal people were never going to do that.
Agent Plugins 1.0 gives everyone a shared way to package it, making other people's expertise much easier to distribute.
ChatGPT, Codex, Cursor, VS Code, GitHub Copilot and Kiro are already on the compatibility list.
Claude isn't. Of course. Claude always has to do things differently, and it ends up making everything harder for the person using it.
I think this is super cool. But every man and his dog is going to be making plugins. watch this space.
Cloudflare built the AI operating system every company is going to want
Cloudflare has been running its own internal AI workspace since May, with thousands of employees using it to create documents, automate jobs and build little apps.
Now it has released Cloudflare OS so other companies can build their own version.
Four things stood out to me:
Your whole team works inside it. You can share skills, chats and finished work like slide decks without passing files and context between a dozen different tools.
It builds connected apps for your team. The apps can use live company data, and multiple teammates can work inside them together.
It runs deterministic workflows. That means simple A-to-B jobs that follow fixed steps. “When this company pays, update the CRM.” No AI thinking needed. The agent builds the workflow once and Cloudflare OS runs it.
You can use any AI model, and the whole thing is open source (the code is public, so companies can change it and make it their own).
+ much more.
I think we’re heading towards a point where every company has some version of this.
I can’t see a future where every employee uses their own ChatGPT or Claude account, with company context scattered across dozens of private chats.
A central company workspace makes far more sense: all the company context and skills built in, access to any model, and one place to build internal apps and share work with the team.
I can see this being the future.
Meta released a Claude Code competitor
Meta dropped Muse Code this week, a coding agent that lives in your terminal (the text-only window developers use to run tools). It goes straight up against Claude Code and Codex.
Under the hood is Muse Spark 1.2, a new 1m token context window model.
On the standard tier, input costs $1.25 per million tokens and output costs $4.25. Your prompts aren't used for training.
Choose the Contributor tier and that drops to $0.10 / $0.20, but Meta can train on your code and prompts. I wouldn't put client work or anything private through that.
I watched this 37-minute review to see what it could do. The review spent most of its time making games, 3D assets and random coding tasks, which is the problem with a lot of model reviews. They barely test the knowledge work we care about.
The landing page was usable, but nothing about it made me want to move over.
For now, it's another player I'll keep an eye on. I can’t see myself using it yet.
Also this week...
Hark is building a browser agent for normal life → Most browser agents have been built around coding and developer work. Handoff focuses on everyday jobs like ordering Uber Eats, shopping, booking restaurants and flights, or messaging candidates on LinkedIn. It’s only a research preview for now, but this feels like the right direction.
Seedance 2.5 can make 30-second videos → Continuous multi-shot scenes and up to 50 references to keep people and products consistent. You can try it now in Krea.
MiniMax released H3 → A downloadable AI video model that makes clips up to 15 seconds with sound. It can reach 2K, but the full-quality version still relies on some of MiniMax’s online tools.
💡 Builder's notes
The newsletter archive
I’m almost at a full year of writing this newsletter, so I wanted the archive to feel like something you’d actually want to browse rather than a boring list of links.
I built a new posts page where every book on the shelf is one of my past newsletters. You can browse through the collection and open any issue from there.
I took a lot of inspiration from Stripe Press, and I love how it turned out.
A Kindle reading skill
I built a skill that finds books online, or takes research papers, guides, long documents and essays I already own, then sends them to my Kindle in the right format.
Before my London to Copenhagen flight, I used it to add all 256 of Paul Graham’s essays and The Count of Monte Cristo to my Kindle.
It even split the essays into themed books and generated proper covers for each one.
the Paul Graham collections it sent to my Kindle
Grab the skill here. Give Claude that link and tell it to download it for you.
Opus 5 is a terrible colleague and a brilliant employee
I’ve been building with Opus 5 all week.. and I can confidently say that I really fucking hate talking to this model.
It’s insufferable. Condescending, backhanded and willing to argue about things it shouldn’t.
Ask one simple question and it will give me 3,000 words of verbal diarrhoea, written in technical language I can barely understand.
But it is very good at doing things.
Kun Chen (@kunchenguid) put it perfectly: AI models are starting to fall into two buckets.
The models you enjoy talking to, and the models that do great work in the background but you can’t stand speaking to directly.
Opus 5 is absolutely the second.
A workflow I’ve been loving this week is using Fable 5 as the orchestrator (the model that keeps the plan straight and delegates the jobs), with Opus 5 as the doer.
Fable stays in the main chat, then spins up Opus 5 subagents (separate AI workers) to do the actual building in the background.
This is literally all I add to the prompt:
Example prompt, with Fable 5 selected:
Please build out [XYZ project] as per the @plan document. Use Opus 5 subagents for all the “doing” work, and keep this main context window for orchestration.
Talk to the model you like. Let it send the hard work to the insufferable nerd in the background.
|
🧰 Tools to try
/wait-what → Your agent just explained something in a 900-word technical mess? Run this and it will try again with more context and simpler language. Basically a “wait, what?” button for AI answers. (free)
QM → YC released the multi-agent system it uses across accounting, legal, events and engineering. It’s designed for several agents working across one company. Interesting, but this is one for the more advanced builders. (open-source)
Comp AI CRM → A CRM built for agents. It reads your team’s inbox, researches companies and contacts, then runs repeatable sales workflows. A very interesting home for an agent managing sales or ops. (open-source)
🥣 Brain food
Alex Hormozi on The Diary of a CEO → I listened to this on the plane. It was phenomenal, as usual. Nuggets in here for everyone.
Alex Hormozi on AI → this image sums it up:
Anywear → Drag clothes from any online store onto yourself and try them on live. Sick.
Make your audio player a cassette → Cool UI idea ill probably borrow (steal)
One thing I wake up feeling oddly grateful for - probably not something you'd expect - is that I found what to work on.
Getting here took 3.5 years of doing work I hated. At the time, it felt like I was stuck.
Looking back, I can see it was part of getting me here: to a place where I get to spend my days building things I actually care about.
Paul Graham's How to Do Great Work captures this perfectly. It's one of my favourite things I've read in a long time. Every time I read it, I smile and think about younger me.
"What should you do if you're young and ambitious but don't know what to work on? What you should not do is drift along passively, assuming the problem will solve itself. You need to take action. But there is no systematic procedure you can follow.
So you need to make yourself a big target for luck, and the way to do that is to be curious. Try lots of things, meet lots of people, read lots of books, ask lots of questions.
When in doubt, optimize for interestingness."
You don't need to have the perfect plan. You just need to stay curious, keep moving, and give luck enough chances to find you.
