I built a control room for my company's skills

plus opus 5's backlash, buzz's 12 harnesses, and a voice clone from 5 seconds of audio

Godmorgen,

Hosted my very first in person meetup this morning in Copenhagen, Denmark!

Was so amazing to meet you guys - to everyone who showed up… thank you.

📌 TL;DR

  • Opus 5 is getting hammered → the timeline reckons it's a regression. I don't think it's that bad, and a lot of it comes down to a setting you can change.

  • Buzz went from 3 agents to 12 → in five days.

  • Perplexity went multiplayer too → Projects turns Computer into a shared workspace, with memory and files that live across your team instead of inside one person's session.

  • Fish Audio clones a voice from 5 seconds → and the API is free for the rest of August, so it costs you nothing to test against your ElevenLabs bill.

  • Builder's notes → how I share skills across my team, the two ways I tried first that broke, and the skills atlas I built to manage the lot.

People really don’t like Opus 5… is it really that bad?


Opus 5 dropped last week, and if you've been anywhere near AI Twitter since, you've seen the pile-on. The general take is that it sucks.

I'll be honest, it's not my favorite model either - but I don't think it's quite as bad as people are making out….

Opus 5 is a very "unpredictable" and goes off on long tangents.

It’s very verbose too, and gives you giant long answers that read like Egyptian hieroglyphics.



I've found the ideal workflow for now… is using Fable as the main conversational model, and have Fable delegating to Opus 5 for execution.

Fable maintains such a high level of understanding. This saves tokens and is a faster neater workflow.

If you are using opus 5, make sure you keep the thinking level to ‘high’ or lower (the more thinking it does, like on extra high & max, the worse it performs)

BUZZ is buzzing…

Quick recap if you missed last week: Buzz is Block's new agent platform, built like a Slack room where your AI agents sit as members alongside your team.

Well, it's moving fast. Really, really fast.

Last week you could pick between Claude Code, Codex or Goose as the agents that run in the background. 3 options.

This week the list is 12, confirmed by Block a few hours later: Buzz Agent, Claude Code, Codex, Goose, Amp, Cursor, Grok Build, Hermes Agent, Kimi Code, Oh My Pi, OpenClaw and OpenCode, plus the option to add your own.

One thing that hasn't changed: there's still no real way to manage your skills or tools (app connectors).

Still too early to move my real work into it. Still absolutely worth watching.

We're moving from single-player agents to multiplayer ones, and Buzz is one of the clearest signs of it.

Perplexity is going multiplayer too

Perplexity are absolutely cooking right now. It feels like they've got a really, really solid product.

They've launched Projects inside Perplexity Computer - Aravind Srinivas described it as turning Computer into a multiplayer agentic operating system for work: memory, files and sessions that persist and are shared across people, instead of trapped inside one person's session.

It's live for everyone too.

Fish Audio can clone a voice from five seconds of audio

Fish Audio launched S2.1 Pro this week, a text-to-speech and voice-cloning model, alongside a $52 million funding round.

Aimed squarely at ElevenLabs and Cartesia, the two names most of you are probably already paying.

It clones a voice from roughly five seconds of audio, gives you word-level control over emotion, intonation and pacing (so you can direct exactly how one word gets delivered instead of regenerating the whole clip and hoping), and supports 83 languages natively.

Their biggest value prop to users is: cost and speed.

They say it runs about twice as fast as Cartesia and costs about 1/6 of ElevenLabs.

Also this week...

  • OpenAI cut its prices hard → Sam Altman knocked 80% off the price of GPT-5.6 Luna. It's just as smart as their top model was five months ago, and now costs about a twelfth as much. That's how fast this stuff is getting cheaper.

OpenAI's own chart, Luna tops the intelligence index at a fraction of the cost per task

  • HeyGen does video podcasts now → give it a document, a link or just an idea, and it hands back a two-host video show with studio scenes, camera cuts and B-roll, in minutes. Clever, but not for me. You can still tell it's AI, and unless you do it tastefully it's a fast way to torch brand trust.

  • Granola landed on Apple Watch → "The best wearable is the one you're already wearing," says Granola's Chris Pedregal, and now you can capture a meeting straight from your wrist.

💡 Builder’s notes

Sharing and managing skills with your team

We all know how good skills are by now.

They're the single most important thing to get right with agents, because building good skills for repeatable work is where the tangible hours saved are;

A proposal skill. An email formatting skill. A creative strategy skill. A Facebook ads reporting skill.

But then you build a brilliant one, and you want the rest of your team using it too. So how do you actually share it?

Package it up as a zip or a .skill file and send it over, and you've immediately got a mess. A bunch of different versions floating around and no source of truth.

Keep it in something like Google Drive or Obsidian instead, and you need fragile hacks like symlinks just to get Claude to read it. Good luck getting that working for your less technical people. You'll end up running a support channel for it.

I looked for months and never found anything anyone had built that was any good.

Then I landed on this: push all of your skills into a GitHub repository and make it a plugin.

Think of GitHub as Google Drive for code, and a repository as a folder inside it.

You turn that repo into a plugin, which is just the way Claude pulls skills out of it, and then everyone on your team installs that one plugin.

It works in both Codex and Claude Code. All your skills live in one place. And when you improve one of them, your whole team gets the update automatically.

But here's where it gets fucking incredible.

I built a web UI to see, visualise and manage my entire organisation's skills.

skills atlas: graph view

It breaks them down by category and shows you which skills call other skills.

skills atlas: library view

One of my favourite bits in the app is the Skills Assistant, where you can just ask it things like "what skills should I use to write the newsletter?" or "what's best for decision making?"

That gets handy once you've built hundreds of them and start losing track, and handy again for anyone new who has no idea what's available to them yet.

The other favourite is the usage section. You can see how often each skill gets used, broken down by team member.

Which means in six months I can look at what nobody has touched once, and start culling the dead weight before the whole stack gets bloated.

I recorded a 1 minute video walkthrough for you :)

🧰 Tools to try

  • Hermes to Buzz profile skill → since you can now use hermes with BUZZ, this is how you actually get your own Hermes profile in there.

  • MiniMax H3 → new video model and the pricing is silly. native 2K with properly synced audio generated in the same pass, up to 15 seconds, for around 13 cents a second. that's about a third of what Seedance costs, and you can feed it images, video and audio together as reference in one go

🥣 Brain food

Next stop London bby, thank you all once again for being here.

feel so privileged to have almost 24,000 nerds tuning each week to hear me ramble :)