G’day mates,
I’m back in the motherland (Australia), for 2 days and then off to Bali for a month.
The big theme for this week is vibe design. Claude design is here, and so is the new Chatgpt image model. Pair them together and thats branding sorted for any project/ business.
📌 TL;DR
Claude Design → Anthropic's Figma competitor, baked straight into Claude. check out these use cases for websites, social content, emails + more.
gpt-image-2 → probably the best AI image model available rn
Workspace Agents → OpenAI's cloud version of a Claude Code agent. Hosted, schedulable, team-shareable.
Kimi K2.6 → Moonshot's open-source 1T param model. Matches Opus 4.7 on coding at a fraction of the cost. Best open option for privacy-sensitive local setups.
Builder's notes → I'm giving away my global CLAUDE.md file, the full Claude Code setup I run in every session.
Babe wake up… Claude design dropped
Anthropic shipped Claude Design, powered by Opus 4.7. It's a dedicated design workspace that lives inside Claude.
Think of it as Anthropic's Figma. Web-only for now, no desktop app yet. Figma stock dropped 4.26% the same day.
Claude Code can already make HTML pages. Claude chat can already spit out PPTX decks. And with a good frontend-design skill, the output looks decent.
Four things are actually different here.
Direct manipulation. You grab an element and drag it where you want. Change padding with a slider, tweak a font in the sidebar. Click-and-edit, exactly like Figma or Canva.
Persistent design system. Upload your brand once (Figma file, GitHub repo, or folder of assets) and Claude builds a proper design system. Brand colors, semantic colors, typography, spacing, logos, light and dark mode. Every project going forward inherits it automatically.
Projects that stay editable. Each project saves as an editable file you can come back to. Weeks later you can open it up, change the hero headline, swap a CTA color, or move a block around.
The handoff. Every project has a "handoff bundle" that packages the design for Claude Code.
I played around with it a bit.
I setup my AI with Remy brand style first, by giving it my website code. I got it to make a pitch deck…
and then I asked it to do some email designs…
My branding is primitive asf so the outputs weren’t as magical as some of the others I’ve seen.
Im going to link dump below my favourite examples ive seen below…
People are already using it for digital e-guides and lead magnets, social media carousels, email designs, and landing pages and websites.
One note: without a proper brand system, AI tools still produce generic-looking slop. If you don't have a brand kit yet, thats a worthy investment. Or get chatgpt to build you one (more on this below).
It's in research preview for Claude Pro ($20/mo), Max, Team, and Enterprise subscribers, at no extra charge. If you already pay Anthropic, you already have it.
OpenAI shipped gpt-image-2
OpenAI released ChatGPT Images 2.0, powered by a new model called gpt-image-2.
Give it a complex prompt and the model actually thinks through it, reviews its own output, and iterates until it lands. Standard mode is free for everyone; thinking mode is paid only.
It's also the first OpenAI image model that can search the web before generating. Ask for a real brand logo, a specific landmark, or a recent public figure, and it pulls reference material from search.
Free on ChatGPT Free, Go, Plus, and Pro. Also via the API.
The big use case this week is branding. People are building logos, moodboards, typography, and brand photography in a single ChatGPT window, then feeding the output straight into Claude Design as a brand kit. That's the vibe design workflow.
Here are some of my fav examples to give you guys an idea of whats possible…
If you make ad creatives, infographics, product shots, or anything with real readable text, this is the new baseline.
Will be testing this vigorously over the next week, and replacing all my nanobanana workflows to test the difference.
OpenAI shipped Workspace Agents
ngl this is a REALLY cool product…
OpenAI just caught up with Anthropic on managed agents, and made them way more user-friendly.
Workspace Agents launched April 22 inside ChatGPT Business. Think of them as the cloud version of a Claude Code agent.
If you've built a dedicated “agent” in Claude Code, you know the pattern.
A folder on your computer with a CLAUDE.md file for the role and context, a set of skills to train it on processes, and some MCPs to connect your tools. It runs on your laptop.
A Workspace Agent is the exact same structure, but with a clean UI and hosted in ChatGPT's cloud.
Custom instructions, skills, MCPs, connected apps (Slack, Gmail, Drive, SharePoint). Runs in the cloud, so your laptop doesn't need to stay on. And you can schedule tasks.
for example; a Friday metrics report generates at 4pm without you opening ChatGPT. or a lead response agent fires every 15 minutes while you sleep.
This is so frustrating for me to see, because OpenAI have pulled ahead here in a lot of ways.
Anthropic has had every piece of this for months; skills, routines, claude code, MCPs, CLAUDE.md files. The Claude ecosystem is incredible, but the problem is it's just not very user friendly.
Take something basic like controlling Claude from your phone. Anthropic has three different products for it: Dispatch, Remote Control, and Channels. Good luck explaining that to a non-technical friend.
Building agents have the same issue. I've got dozens of them in Claude Code living on my laptop. The second I shut the lid to go run errands, they stop.
This is almost entirely due to the fact that these claude agents are local agents; running on your actual computer with access to your files.
Keeping them running 24/7 means either a dedicated machine at home (a Mac mini on a shelf) or a cloud setup wired through a GitHub repo. Or claude managed agents handles this now, but its a steep learning curve even for me!
And sharing agents with your team is borderline impossible in Claude Code.
Workspace Agents sidestep all of that. You build in a browser, share them with your team, and it runs in the cloud without you babysitting it.
I'll still use Claude for everything personally. The models are that good, AND claude have had all these features like 3-6 months before OpenAI, which is inexcusable.
Maybe the price to pay for shipping lots of features fast, is confusing your beginner users 🤷♂️
But this is a legit win for OpenAI, and it's a bit silly Anthropic hasn't made their ecosystem easier for general users.
Kimi K2.6
Moonshot AI (a Chinese AI lab) dropped Kimi K2.6 on April 20, and it's 100% open source. You can download the whole thing, run it on your own hardware, and use it commercially. No restrictions.
On coding, it's close to Claude Opus 4.7 and a lot cheaper per token. It's the best open model that's ever shipped, full stop.
Closed models like Opus 4.7, GPT-5.5, and Gemini 3.1 Pro still edge it on reasoning and long-context work. It also overthinks simple problems, which burns extra tokens and eats into the cost advantage. Most community reports land around "it's 85% of Opus, not better, just cheaper."
Quick refresher on "local models." Claude, ChatGPT, and Gemini all live on someone else's servers. Every prompt you send leaves your machine, gets processed on theirs, and comes back. A local model flips that. You run it yourself, and your data never leaves your network.
If you want to run a Hermes agent or OpenClaw on a budget, K2.6 is the best cheapest option right now.
Reality check on running it locally. Even a maxed-out Mac Mini can't load it (roughly 4x too little memory). You're looking at around $20K of gear to run K2.6 on your own hardware. Think two of Apple's most expensive desktops wired together, minimum.
So most people will end up using Kimi through Moonshot's API or a hosted provider like OpenRouter. Still cheaper than Claude, but not private, because your prompts run on someone else's servers.
For regulated businesses (cannabis, peptides, finance, healthcare), $20K of hardware is the price of keeping customer info and IP out of someone else's cloud.
📝 Builders notes
I've been getting a few questions about my setup recently, so I've decided to share my global CLAUDE.md file this week.
My global CLAUDE.md file for all you nosey bastards who keep asking.
This is the full Claude Code instructions I load into every session, sanitized. The bits worth stealing:
Plan mode as the default. For anything non-trivial (3+ steps or an architectural call), Claude plans before building. If something goes sideways mid-task, it stops and re-plans instead of pushing through. Kills the "Claude runs off and does the wrong thing" risk before it starts.
Subagent-first orchestration. The main context window is sacred. It's for decisions and talking to me. Anything grunt work (reading 3+ files, broad searches, running tests, producing intermediate analysis) gets delegated to subagents, and their results flow back as summaries, not raw output. This is the single biggest unlock for long-session work, because it stops the main context from getting polluted with tool output.
Obsidian as the second brain. My vault is the single source of truth for every bit of business context (brand guidelines, strategy docs, meeting notes, client info, decisions). Every workspace loads the relevant bits via @imports at session start, so context flows in automatically.
Verification before done. Nothing is complete until it's proven to work. Run the tests, check the logs, demonstrate correctness. "Would a staff engineer approve this?" is the internal check before anything gets marked done. Kills the "ship it and see" instinct.
Boil the Ocean. The standard isn't "good enough," it's "holy shit, that's done." Stole this from Garry Tan
🧰 Tools to try
Claude Managed Agents Memory → now in public beta. Your agents learn across sessions instead of starting from scratch every time.
An open-source version of Claude Managed Agents → lets you use any model, self-hosted, with credential isolation.
🥣 Brain food
AJ Tracey x Jorja Smith Live Lounge → had this one on repeat all week.
Intro to Claude Design → my friend Adrien made the perfect short tutorial. Watch this if you want to actually learn to use it.
Claude Design in action → some more cool use-cases
Top 10 Claude Code Skills, Plugins & CLIs → solid refresh on what's worth installing right now.
What is Claude Managed Agents? → 3-min Anthropic explainer to get your creative juices flowing
I’ve been building solutions for Comfrt all week, and unfortunately cant disclose too much on that front.
Ive also been completely nerfed. I unplugged my Mac Mini and brought it with me in my suitcase to America, then to Bali. the plan was to plug it in at the hotel and let my agents run 24/7 in the background while I'm out doing my thing.
im an absolute vegetable because I didn't bring a keyboard or a monitor, so even though it's sitting there plugged into the hotel wall, I can't log in to connect it to the wifi.
So I've been carrying around this useless brick in my suitcase for two weeks, back to running Claude Code on my laptop like it's the Stone Age.
will have that fixed next week,
