should you actually use Grok Bot?

whats up ladies and gents,

bit of a late one this week

just got back from a ski weekend in the Victoria's High Country, weather was insane ☀️

📌 TL;DR

• SpaceXAI's Grok Bot gives every agent its own computer → it sits in the cloud with its own Chrome, terminal and logins, so your agents keep working with your laptop shut
• Grok 4.6 is frontier level for a fraction of the cost → best at the boring multi-step stuff like research and drafting, worst left alone to run
• Claude can now use any website for you → it reads the page and clicks around the way you would, even on sites with no AI in them at all
• Claude now watermarks what it writes → an invisible marker in everything the newest models produce, and no way to turn it off
• Opus 5 is genuinely terrible, and here's what I use instead → one line: Fable 5 as the brain, Opus 5 as the hands

---

SpaceXAI launched Grok Bot, the simplest AI agent platform yet

SpaceXAI launched Grok Bot in early beta. I played around with it on a trial plan and it's insane.

It's like OpenClaw and Claude Code if it was extremely simple to use and actually had the features people want.

Creating your own bot and connecting your tools takes minutes.

My favourite part: unlike ChatGPT or Claude Code, every bot gets its own computer in the cloud. Its own Chrome, file explorer, terminal and logins.

So it keeps working even if your laptop's shut.

There's an iPhone app too, so you can pick the same conversation up wherever you are.

You can also show a bot how to do a job once. It saves it as a routine and runs it on a schedule from then on.

And the bots talk to each other. One can hand a job to another without you in the middle.

Now here's the catch: it's pricey.

It's only on the top tier plans, like the $200 a month Cursor Ultra one, which, I'm gonna be honest, I'm not getting after already having a Codex and a Claude subscription.

I'll just keep using Claude Code for now. It doesn't seem worth switching yet.

But if they lower it, I'm on this ASAP.

This could be the thing that finally gets regular people using AI agents.

---

Grok 4.6 just got released

xAI just dropped Grok 4.6 on 12 August.

It's on the frontline with Fable 5 and GPT-5.6 Sol, while being extremely cheap.

It's genuinely best at the boring multi-step stuff: research, drafting, building out documents.

It's worst at being left alone to run.

To put the cost in perspective: about $2 to feed it eight novels' worth of text, and about $6 for it to write eight novels back. That's a fifth of what GPT-5.6 Sol charges for the writing half, and an eighth of Fable 5.

Who this actually matters to: if you're on a normal monthly plan, you never see any of these numbers. It just means your usage stretches a lot further before you hit a wall.

If you're building something that talks to the AI directly, this is literally your bill.

---

Claude can now use any website for you

Okay so Anthropic updated Claude in Chrome on Tuesday, and it's a bigger deal than the announcement made it sound.

It's just Claude sitting in a panel next to whatever site you're looking at, except now it can go and actually do things on that site.

The useful part is that it works on sites with no AI in them at all. Internal dashboards, old vendor portals, that clunky system your supplier makes you use.

It reads the page and clicks around the way you would, while you're already logged in, and it still asks first before anything you can't take back.

---

Also this week...

Auto mode becomes the default in Claude Code → from today it stops asking your permission before each action and has a second AI check the risky ones instead. Paid Claude Code plans only, and only if you never picked your own setting. Shift+Tab puts the asking back.

Meta built a Claude Code competitor → a coding agent that runs in your terminal, going after Claude Code and Codex. It's powered by their new model, Muse Spark 1.2. Nothing here would make me switch.

Claude now watermarks what it writes → Anthropic is hiding an invisible marker inside Claude's text so software can tell it came from an AI.

Every Claude model launched from August 2 onwards ships with it, worldwide, with no way to turn it off. They're working on adding it to the older models too.

It survives copying and pasting. Heavy editing wears it down.

No detector has been released yet, so nothing can actually check it today, but Anthropic says that's coming.

Worth knowing before you paste Claude's work into anything that gets checked. And it still counts when all you asked for was a tidy-up of your own writing.

---

💡 Builder's notes

Opus 5 is genuinely terrible

After using this model for a couple weeks now i Definitely say This model fucking sucks.

On paper it's really good, but it honestly depends on how you use it.

It is the one and only Claude model I've used that has been absolutely insufferable.

It makes so many mistakes every time I use it. I'll ask it, "hey, why did you make that mistake?" and it'll say, "I have no idea."

There's no excuse for that.

like wtf?

So I've honestly just given up and started using Fable 5 on low effort.

I'm only using Opus 5 as a subagent working directly under Fable 5, because it is genuinely good as a work agent.

This is the rough prompt I use:

Please build x. Use Claude Fable 5 as the orchestrator brain,
        and use Opus 5 as sub-agents to do the work.

sorry for the rant lmao

Making Claude write in your voice

There's a skill called Humanizer. It's built off the Wikipedia article on signs of AI writing, so it knows 33 patterns that give AI away and strips them all out.

Problem is, once they're gone your writing doesn't sound like you. It sounds like nobody.

So I tweaked it. I added a references folder, one file per kind of writing, with how I write at the top and ten real things I've written at the bottom.

I built the email one first. Dropped in ten emails I'd actually sent, and now anything it writes for me opens with "Hey" and signs off "Cheers", because that's what I do.

You can build one for whatever you want. Marketing copy, content, essays, proposals. Same document, different examples.

Here's the prompt that builds it: Make Claude Write Exactly Like You

---

🧰 Tools to try

• /wait-what → when your AI buries you in 900 words of jargon, this makes it explain again properly.
• Airship → design your interface visually instead of describing it to Claude in words.
• GlassQuota → see how much you've got left on Codex, Gemini and Claude from your Mac menu bar.

---

💎 Cool finds

• Motion designers are so cooked → this AI agent made the whole video in 37 minutes, and only needed four changes.
• A guy built a time zone app in a few hours with Codex → tells you what time it is for everyone you work with. What a time to be alive.
• An animated receipt when someone subscribes → someone made signing up print you an actual receipt. Completely pointless, and the exact kind of detail people screenshot (might have to steal this one).

---

thats it beautiful people

but to be honest i’m not too happy with the state of my content, and content in the AI space generally.

Remy Gaskell, another creator in my niche, made a video saying AI content has become a giant cesspit of reused AI slop.

it really opened my eyes, ive even caught myself doing it, making agents to research content thats already working.

so im taking a step back and trying to be as original as i can be.

thanks guys.

PS: reply if you have any questions, happy to answer anything :)