I've never enjoyed AI this much
My new favourite way to work. Plus Claude Code mods, Grok Bot and Nano Banana 2.1
Remy Gaskell
October 10, 2026
|
📌 TL;DR
Claude Code mods → describe a feature and Claude builds it into itself.
Grok Bot → use any model, Team Bots and X, and why it's winning at work automation.
Claude in Google Docs → what is going on gemini
Nano Banana 2.1 → Google's new image model: how does it compare against GPTs?
Haiku 5.5 → Claude's cheapest, fastest model, built to be Opus's sub-agent.
Builder's notes → my new favourite way to work with AI, why I build cloud-first, and the browser I give my agents.
Claude Code mods: describe a feature, and Claude builds it into itself
Anthropic launched mods for Claude Code, and it's one of the coolest releases I've seen in a while.
A mod is a small add-on that changes how Claude Code looks and how it works.
You already know skills (a handbook that tells Claude how to do a job) and connectors (which plug Claude into your other apps)…
a mod is different because it changes Claude Code itself. It can add buttons, side panels and status bars, or catch what Claude is about to do and change or stop it.
This RoboNuggets video is the best walkthrough I've seen. here are some examples of mods:
Agent progress. A progress bar above your prompt, plus a live panel of every sub-agent and how far along each one is.
 |
Skins. Completely change how Claude Code looks, right down to the desktop app. He gave Claude a colour palette and asked it to build a skin from it, which is freaking cool.
 |
 |
the finished skin in use |
Reflect. When you correct Claude ("no, I wish we'd done it this way"), a pop-up asks whether to save that as a permanent rule in your instructions file. Click save, and Claude writes the rule in for you.
 |
+ so many more
Claude Code now ships with a built-in skill for building mods, so to build your own mod just describe what you want ("make a mod that shows my usage above the prompt"), approve it, and it shows up straight away.
I can see this becoming a trap, where you build a pile of useless mods for the sake of it, feel productive, and nothing in your business actually moves.
So tread carefully. or have some fun with it, but call it fun time instead of convincing yourself it's meaningful work.
Grok Bot: the team has been busy
Quick refresher if you haven't used it: Grok Bot is xAI's always-on AI agent. you can spin up different bots with their own persona, role, and responsibilities, with each bot having its own dedicated virtual computer.
 |
latest grok bot releases you should know about:
It'll pick the best model for each job. On 7 October, Elon announced that Grok Bot will "use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs."
Simple questions go to small, fast models, and complex ones go to big models. You don't get a model picker, and there's no date yet.
Team Bots. One shared agent your whole team can use. The owner sets it up once with its skills, files and app connections, then teammates chat to it in the Grok Bot app or tag it in Slack. It uses each person's own logins, and every chat stays private. Team Bots are in public beta on Teams and Enterprise plans.
It plugs straight into X. Connect your X account and your bot can search posts, read your timeline, check your mentions and sum up what's happening. You can even tag @bot under a post to hand it over as a task.
This is insane, because X is the lifeline of the internet whether you like it or not.
Before, you needed a bunch of hacks, APIs and an expensive X developer account.
You can call it. Talk to Grok Bot out loud and it goes off and does things while you talk. I can imagine a 20-minute car ride where you're chatting to it and working through all your emails.
Grok bot for work automation
imo AI products fall into two clear categories…
work automation
or
work amplification
Work amplification is AI that makes you better at the work you already do. You're still in the driver's seat, it just makes you faster and sharper.
Work automation is AI that does the work that shouldn't need you at all. It runs in the background, often for the whole team.
Or put simply: Claude Code makes the work you do better. Grok Bot does the work that shouldn't need you.
And Grok Bot is starting to look like one of the strongest automation options on the market, if not the strongest.
Mainly because these updates fix my two biggest gripes:
No model choice. I started building Grok Bots to help with parts of my newsletter scripting, and the Grok model's copywriting wasn't amazing. I had no way to swap in a model that writes better. Model routing should sort that out.
No team bots, the bigger gripe. For an agent to actually automate work, your team needs to use it too. Picture a sales agent loaded with your proposal skills and CRM data, that three people on your sales team can all talk to, including the non-technical ones in Slack. That's exactly what Team Bots allow.
I think Elon has built one of the strongest AI products on the market. This space is worth watching closely.
For amplification, Claude Code is still my tool. I tried building things in Grok Bot and couldn't get my head around it. The interface wasn't the right place for me to build in.
Codex & Cursor are solid amplification options too (and cursor links up with Grok Bot)
So it isn't one or the other. Each one fills a clear role, and they can work together.
Claude now works inside Google Docs, Sheets and Slides
Anthropic launched a Claude add-on for Google Docs, Sheets and Slides. It opens Claude in a sidebar next to the file you've got open, so it can work on that file directly.
Honestly, I thought this already existed. Apparently not. Unless Anthropic is sneakily re-releasing the same feature twice and hoping we won't notice. But cool, I guess.
 |
This puts Claude up against Gemini inside Google's own apps.
Nano Banana 2.1: better, but not the best at everything
Google released Nano Banana 2.1, its new model for making and editing images
It's a real upgrade on Google's older image models. But it doesn't beat OpenAI's GPT Image 2.5 across the board.
The most useful test I found was from Wanderson Jackson, who ran 13 rounds through both models with the same prompt and the same product each time.
Nano Banana 2.1 won editorial portraits and 3D, Pixar-style characters, and it had the better layout and overall look.
 |
GPT Image 2.5 won candid street photos, UGC-style selfies, cinematic stills and product ads.
 |
The ads are the big one. Nano Banana reshaped the product (it rounded off a rectangular bottle), and no brand will accept a distorted product.
 |
Nanobanana 2.1 is "faster, more literal, and prints the numbers you ask for", while GPT "holds brand details tighter and owns lo-fi realism".
So:
Use Nano Banana 2.1 for thumbnails, posters, infographics, labels and detailed design briefs.
Use GPT Image 2.5 for UGC, cinematic shots and product ads.
Claude Haiku 5.5 completes the 5.5 family
Anthropic released Haiku 5.5 and it's the smallest, fastest and cheapest model in the new Claude 5.5 family.
Anthropic pitches it as a sub-agent for Opus and Sonnet: the bigger model plans the work, then hands smaller pieces to Haiku to knock out quickly and cheaply. It's also built for high-volume, simple jobs like summaries, sorting and live support.
 |
If you pay per use, it's about 75% cheaper than the last Haiku on average, at $0.10 per million input tokens and $0.50 per million output tokens for normal-length prompts.
Also this week...
Google Docs and Drive now open Markdown files natively → which is a massive one. Docs now opens and co-edits .md files without converting them, and Drive shows them as formatted previews. So files like your AGENTS.md, CLAUDE.md and skill files can be shared and edited in Docs. It's free for every Google account and rolling out over the next couple of weeks.
ChatGPT can now take your meeting notes → a new Meetings plugin in the ChatGPT Mac app takes notes on calls and in-person conversations, with no bot joining the call, then gives you a summary and next steps.
Granola now snapshots shared screens → Granola can grab the slides, charts and designs someone shares in a meeting and use them in your AI notes. It's Mac only, and for now it works with screens shared in Microsoft Teams.
💡 Builder's notes
1. Claude Projects + Opus 5.5 is my new way of working
I've never enjoyed working with AI as much as I have in the last week with Opus 5.5 and the new Claude Projects feature.
Opus 5.5 is the most amazing model I've ever used. For full transparency, I'd moved over to Codex for a while because Opus 5 was so bad. 5.5 brought me back.
The new version of projects (projects v2) is the bigger deal though, and it's hard to picture until you see it.
So here's my Newsletter project:
 |
Left: one project per department of my business. Growth, Marketing, Content, The AI Course, Newsletter, Finance.
Middle: the main chat. It's like a chief of staff agent. I hand it tasks and it spins up a separate thread for each one. Those threads are real, fresh Claude sessions you can jump into and talk to (not sub-agents).
Right: a Kanban-style board of every thread. What's working, what's waiting on me, what's idle and what's resolved.
The Library tab shows every file and folder the agents have created for themselves, all in the cloud on their own little virtual machine.
 |
Then in Routines you can set up scheduled tasks inside that project. My weekly AI research skill now runs every Wednesday, right inside the Newsletter project.
It's all in the cloud, so it runs even when my laptop's shut, which I think is just fucking phenomenal.
 |
The same projects and threads open in the Claude iPhone app, and they're the same across Claude Code and Cowork.
It's pretty much Claude's answer to Grok Bot and OpenAI's Dots.
I can't even imagine working any other way now.
Heads up, Projects is still in beta, so not everyone has it yet.
2. Build cloud-first
My rule of thumb: the world is moving to everything in the cloud, so build accordingly.
Most of our agents still run locally on our own computers, like my OS folder. That means:
shut your laptop or catch a flight and it stops (scheduled tasks included)
you can't use the same setup from your phone or a second computer without hacky workarounds
a teammate can't jump into your session and work on it with you
We went from chatbots, to agents, to agents running on our computers. Now we're getting always-on agents in the cloud with their own computers. Claude Projects, Grok Bot, OpenAI's Dots.
Think of it like Word docs going to Google Docs. You used to write the doc on your computer and email it over for review. Now that seems stupid. Everyone works on the same doc, in the cloud.
The fundamentals don't change. Context, tools and skills are still how you set up an agent. Only where they live is changing.
Your instructions file, skills and context files are the real assets. Most of the other stuff in your OS folder is scratch work.
My setup right now:
Tools connect through Composio.
Skills live in a plugin backed by Notion and GitHub.
Finished work and shared context go into Notion, which every agent can read and write.
My local folder is just a sandbox for messy drafts.
It's an awkward transition. None of these cloud agent products feel fully finished yet, and some still make you set up GitHub repositories and environments (developer land none of us want to go down). Projects is the closest I've seen to removing that step.
But if you're trying to build as future-proof as possible, always opt for the cloud-first option where you can.
3. Browserbase: give your agent its own browser
A few of my skills used to drive the browser on my Mac, clicking around in tools that have no API or MCP (the connector that lets Claude use another app). The moment I moved to Projects and my phone, they broke.
Say I'm at the airport, my computer's checked in, and I want to post a video. Claude goes "great, I'll create this ManyChat flow for you. Hang on, your computer's off. I can't use your browser." And then it's game over.
Browserbase fixes that. It gives your agent its own little cloud browser to click around in. I've got it connected to Claude through Composio and I'm on the $20 a month plan. It's worth it.
Two ways I use it:
ManyChat before every reel. ManyChat runs my "comment this word and I'll send you X" automations. It's how I grew my email list to about 36,000 people. Every reel needs its own flow set up, with the trigger pointed at that video. ManyChat doesn't have an MCP, which is just stupid. So now I ask for it from bed and the agent does it in its cloud browser.
YouTube title and thumbnail split tests. My upload skill tests 3 titles and 3 thumbnails. The YouTube connection can upload the video but can't set up the split test, so Browserbase does that bit.
On logins: Browserbase saves your login session (the cookie), not your username and password. You log in once and it stays logged in, the same way LinkedIn does in your browser. You can opt out. Weigh up the trade-off yourself.
4. Make Claude feel less Claude-y
Small one. I switched Claude's interface font from Anthropic Sans to System.
I love the Claude branding, but the app felt very Claude-y, and I wanted something that felt more modern and neutral.
System makes the app feel a little better, and a little closer to Codex, whose interface I much prefer. You'll find it in Claude's settings under Appearance > Interface font.
 |
|
🧰 Tools to try
Grok Bot's Shopify connector → Connect your Shopify store to Grok Bot, then ask it to check orders, track inventory and keep your product listings up to date (just announced)
🥣 Brain food
speedrunning an entire startup with AI in 60 minutes lol → Thomas and I build a whole startup with AI, in 60 minutes.
pov: you're building an AI startup with friends in Bali → my new vlog.
rick ruben → fascinating listen
thanks for tuning in folks :)
and welcome to all you new comers, theres 37,000 of us now
