Cmd+AI

Kabarza and Samuel catch up after a birthday week, talk about building too many side projects and why repeatable work is worth systematising, then go through Apple's cheaper MacBook Neo and the M5 MacBook Pros, Webflow's new component canvas, and OpenAI's GPT-5.4. Samuel sets GPT-5.4 a redesign task with a front-end design skill, and the results lead to a wider argument about how much design most products really need.

Along the way they try web haptics on their phones, look at Alibaba's Qwen 3.5 running locally, weigh up whether Perplexity Computer is worth $200 a month, and cover Gemini 3.1 Flash-Lite, Claude Code's voice mode and Wispr Flow, and Figma's new slots. They finish on Anthropic being labelled a supply chain risk by the US Department of War, and what OpenAI's move to take the contract says about both companies.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel flies solo this week while Kabarza celebrates his birthday. He opens with how OpenClaw now produces most of the show, why he is resisting the urge to pile more work on with AI, and his first go with Codex agent swarms, before turning to Andrej Karpathy buying a Mac mini to tinker with "claws" and Anthropic's claim that DeepSeek, Moonshot and MiniMax ran industrial-scale distillation attacks on Claude, including why the internet was unsympathetic and whether the numbers add up.

He then answers viewer comments on the Figma MCP's new code-to-design tool, looks at Figma's incoming AI credit limits, and asks why Claude Code Security knocked cybersecurity stocks. The second half covers background and remote agents from Cursor and Claude Code, agent teams and orchestration, how he uses OpenClaw skills day to day, and Anthropic's refusal to drop its limits on mass domestic surveillance and fully autonomous weapons for the US Department of War, before a closing word on next week's Apple announcements.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel start with a catch-up: why AI coding stopped feeling scary, why you should begin with a problem you already have rather than the technology, and an honest look at whether content creators can make a living. They then cover Google's Pomelli product-photography tool, Peter Steinberger joining OpenAI and what his move says about Europe's attitude to AI, Claude Sonnet 4.6, Gemini 3.1 Pro's animation output, and a claim that 4% of GitHub commits now come from Claude Code.

The main segment is Figma's Code to Canvas, which sends a vibe-coded app from Claude Code into Figma and back again. Kabarza explains how it changes the hand-off between designers and developers, and the pair discuss how agencies should price app work now. They finish with Anthropic's new rules banning Claude subscription tokens in third-party tools such as OpenClaw, and whether its walled-garden approach is costing it goodwill with developers.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel open with Jony Ive's design for Ferrari's first electric car, which they find beautifully detailed but not very Ferrari, then look at Seedance 2.0 and what near-cinema-quality AI video means for indie filmmakers. They also discuss the leaked OpenAI "Dime" ad and whether an audio-first device makes sense. On the coding side, they cover Anthropic's Opus 4.6 fast mode at six times the price, OpenAI's GPT-5.3 Codex Spark running on Cerebras, and why Kabarza has moved most of his work to the €20 Codex plan.

Kabarza explains why he feels misled by Higgsfield's ads, and Samuel takes a careful look at the #keep4o movement and what it says about people relying on agreeable chatbots. Samuel argues that agent orchestration is the skill to learn in 2026 and uses Warp's new Oz cloud agents as an example. Kabarza then tests Webflow's official Claude connector live, publishing a CMS item and having it build a testimonials page. The episode ends with Framer's new server API for the CMS and how tools like Notion or Airtable could handle editorial workflows.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel is back from a business trip to Sofia and Kabarza admits he is sitting on top of Mount Stupid, already offering clients apps he has vibe-coded. That turns into a frank argument about learning on a client's time, security, GDPR and why the Moltbook founder's "I didn't write one line of code" boast went so badly. Then the week's news: Opus 4.6 and GPT 5.3 Codex land 25 minutes apart, Claude Code's agent teams, and Kabarza compares OpenAI's new Codex desktop app with the code tab in the Claude desktop app.

Kabarza walks through Webflow's new AI site builder and its canvas view, what the Flowkit style guide means for Webflow developers, and why the canvas matters more than the AI. Samuel explains how Moltbook fell apart, from an exposed database to missing rate limits and crypto spam. They also cover usage limits and replacing paid tools with your own, Anthropic's ads poking fun at ChatGPT and Sam Altman's reply, Figma's new image-to-vector feature, malware in a popular OpenClaw skill, and a first look at the cocktail app they plan to build properly.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel and Kabarza spend most of this episode on OpenClaw (formerly Clawdbot and Moltbot). Samuel explains how he runs it as a cautious assistant: hardware, VPS options, a SOUL.md file with rules it can't override, and building trust the way you would with a new hire. They also cover prompt injection and Moltbook, a social network where AI agents post to each other. Before that, they compare Pepsi's well-received ad with Coca-Cola's AI ad and talk about how brands jumped on the penguin meme.

Kabarza tries Variant AI for generating design ideas. Samuel gets Kimi K2.5 to build a deliberately over-the-top landing page and connects Pencil.dev to Claude Code. They then cover Webflow's new community hub, Claude Code's shared task lists, slash commands being merged into skills, Vercel's finding that AGENTS.md beats skills, MCP Apps and Google's Project Genie, including a viral Genie clip that turned out to be fake.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza explains how he planned a small room in Blender without knowing the software. He measured the room, gave Claude Code a Figma sketch and the Blender MCP, and found that asking ChatGPT for a detailed JSON prompt fixed a bed that kept exploding. Samuel describes Remotion and its new agent skill for turning React designs into MP4s. The hosts then look at Vercel's skills.sh package and other skill directories, and Samuel explains how skills differ from slash commands and agents. After that they discuss Cloudflare's acquisition of Astro and what it could mean for Webflow.

The second half starts with OpenAI's ad test in ChatGPT: personalised versus irrelevant ads, Google's leaked "Selfish Ledger" concept, and why "sponsored" labels aren't clear enough. They cover Apple's reported Siri chatbot built on Gemini, and Kabarza explains how ChatGPT refused to help him find alternative words for a text about Iranian protesters. That leads to a discussion of Anthropic's published constitution and who gets to decide what is ethical.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel open with a speed-reading app Kabarza vibe-coded in Cursor, then talk about where designers fit in AI-assisted coding and plan a live session to build a cocktail-ordering app. Samuel explains the Ralph Wiggum technique: run your coding agent in a loop with fresh context each time and keep its progress in an external file. He argues that orchestration now matters more than the choice of model or tool. They follow this with Cursor's experiment of letting agents build a browser over a week, and Tool Search in Claude Code, which loads MCP servers only when they're needed.

They also discuss Apple choosing Google's Gemini to power Siri, Claude Cowork as "Claude Code for normies", and Anthropic blocking Claude subscriptions in OpenCode. OpenAI's purchase of Torch leads to a longer talk about health advice, citing sources and approximate truth. The episode closes with "code is cheap, software isn't", Apple's new creative app subscription, Webflow's focus on optimising sites for AI, and how people will learn as tutorials dry up.

Chapters

Watch the full live stream and more at cmdaishow.com.

Samuel and Kabarza are back after a fortnight off. They catch up on Kabarza's move to Payload CMS, his interactive Christmas card for clients and a Rubik's cube race that turns into a lesson on God's number. Kabarza's site of the week is the Getty's 3D Persepolis experience, which he picked in solidarity with protesters in Iran. He praises its accessibility settings and its unusually complete slider.

The main discussion is the Tailwind paradox: usage has never been higher, yet revenue has fallen far enough that the company laid off most of its engineers, because AI writes the classes and nobody visits the docs any more. They go on to Andrej Karpathy's post about feeling behind as a programmer, work through Boris Cherny's tips on how he uses Claude Code, and look at Cursor's dynamic context discovery for MCP tools. The episode closes with MKBHD's collaboration with Epic Spaceman on how small transistors have become.

Chapters

Watch the full live stream and more at cmdaishow.com.

Kabarza and Samuel open their Christmas episode with a WebGL Christmas card Kabarza vibe-coded for his clients, and Samuel's frustrating first week with the Work Louder Creator Micro, which turns into a debate about Stream Decks and why companies should answer feedback where their users are. In the news, Samuel explains where Gemini 3 Flash sits in Google's line-up, its high hallucination rate and the new stateful Interactions API, then runs through a busy week of Claude Code updates, from browser control to prompt suggestions and the plugin marketplace.

Kabarza tests OpenAI's new image model against Nano Banana Pro, and both ask whether AI images are good enough to put in front of clients, with Samuel sharing his own AI photo shoot from 2022. They also cover Replit's write-up on its testing agent, the backlash over GitHub charging for self-hosted Actions runners, Framer's $100,000 creator awards, and the Wall Street Journal's visit to Anthropic's AI-run vending machine.

Chapters

Watch the full live stream and more at cmdaishow.com.