MaibornWolff

Creator Media Empires: Hailey Bieber and the Rhode To e.l.f Beauty Hailey Bieber and Rhode are now the single biggest case study in creator-brand exits. Full stop. She launched Rhode in June 2022 with three products: the Peptide Glazing Fluid, Barrier Restore Cream, and Peptide Lip Treatment. Minimalist line. ✔️ No retail partners ✔️ No Amazon storefront ✔️ DTC only, sold through rhodeskin.com . The brand did $212 million in net sales in the twelve months ending March 2025. It doubled its consumer base in a single year. In May 2025, e.l.f. Beauty acquired Rhode for $1 billion. $600 million in cash, $200 million in e.l.f. stock, and up to $200 million in performance-based earnouts over three years. The deal closed in August 2025. Rhode launched at Sephora US in September 2025 and generated $10 million in sales in its first two days. It captured 2% of total Sephora skincare sales within a week and landed in the top ten brands by revenue. International expansion into Sephora UK followed in November. 💡 💡 For context on speed: Charlotte Tilbury sold to Puig for $1.5 billion after decades. Byredo sold to Puig for roughly $1 billion after 16 years. 💡 💡 The P&L was unusually tight. Cost of goods near 19%, marketing spend around 11%, EBITDA margins in the mid-30s. That marketing efficiency exists because Hailey's 50+ million Instagram followers are a permanent, owned acquisition channel. Only 15% of Rhode's earned media value in 2024 came from posts that mentioned Bieber by name. The rest came from a wider community of micro-influencers and everyday users she cultivated. Hailey moved from being the face of the brand to the founder of an institution. She focused on skin science over flashy trends. She stayed as Chief Creative Officer and Head of Innovation after the acquisition. This is a brand that a large conglomerate already wanted to buy, and did buy, for a billion dollars. Her restraint was her moat.

The AI productivity conversation is not a personal-productivity problem (anymore) - it's more about the challenge to fan it out to the team. Spent last week getting Claude Desktop into the hands of our non-engineers so they could edit our CMS without filing a ticket. Claude can do it on my machine with a bit of handholding, no problemo. What's missing is the layer between "I have it working on my laptop" and "seven colleagues on heterogeneous machines have the same setup without me sitting next to them." 🫣 I shipped a private marketplace on GitHub and a plugin that bundles the MCP servers, the shared tone-of-voice guideline, and the CMS-specific skills. One install command, version bumps in one place, everyone on the same setup. That worked.... after having to debug the second and third setups that were different for my colleagues 😅 What it makes obvious: to make non-developers agentic, you still need a developer. Someone has to author the plugin, decide which MCPs to bundle, maintain the marketplace, write the skills. Once bootstrapped, it's permanent. Most of the tooling discourse is still about how I set up my own editor, my own prompts, my own context. The actual problem is how to scale beyond their own setup - independent of technical knowledge. This is our approach for hyground as well: Using AI in production safely, scaling beyond "it works on my machine". Because that is actually the not so easy part.

We are hiring Lead AI Engineers. If Industrial AI in manufacturing, particularly MedTech, is your thing, reach out to me directly. Different locations possible (check out where we're located at MW). I'd love to talk to people who are truly excited about the topic - random one-liner requests unfortunately won't get a reply 🙏🏻 Looking forward to hearing from you!

That’s a wrap! I just finished my in-person iSAQB® – International Software Architecture Qualification Board Advanced seminar “Web Architectures” here in Munich, organized by Albion Academy GmbH and tecnovy . As always, it was a lot of fun, with many interesting discussions. A heartfelt thank you to all participants! It’s great to see that people still put in the work and hone their skills. “Power is nothing without control” was an old slogan from a Pirelli tires campaign. These days, “AI is nothing without understanding” might be more to the point.

Claude is not impressed with me checking out local models

The programme of Agile Testing Days | Nov. 16-19, 2026 has been published, and I'm thrilled to announce that I'm co-hosting two sessions this year: "Come to the git side, we have Pull Requests" together with my dear friend Andrea Jensen . A 3 hour workshop where we will explore how testers and quality people can contribute to and benefir from pull/ merge requests. No technical knowledge required and we will get our hands dirty with some real world inpired exercises. You will leave with quite some new tools that will up your game! "Kanban and the art of woodworking" together with another dear friend, Søren Wassard . We will build wooden toys, using real physical tools (saws, hammers and others) to build woodedn toys, and we will use Kanban to plan and execute the work. No worries if you no nothing about wood working and/ or kanban, we have you covered! There's plenty more awesome sessions at this years #agileTD , s join us there, and ping me for a discount code that saves you an extra 15% on top of the early bird price! 🤑

🚀 Excited to announce I'll be speaking at Global Azure Tunisia 2026 in just 3 days! 🎤 From Code to Cloud: Deploying Your First App on Azure Container Apps 🕥 Saturday, April 18 · 9:45 AM · 45 min 📍 Mediterranean School of Business, Lac2, Tunis We're going beyond slides. Expect a live demo packed with real-world scenarios: auto-scaling under pressure, managing replicas, and rolling out revisions in Azure Container Apps. Whether you're a developer, student, or IT pro, this session will take you from "what even is the cloud?" to deploying a production-ready containerized app on Azure. 🐳☁️ 🎫 Free registration: https://lnkd.in/djmiAUzr ℹ️ More info: https://lnkd.in/dbRRipx4 See you there! 🙌 #GlobalAzure #Azure #AzureContainerApps #Cloud #Tunisia #TechCommunity

𝗬𝗼𝘂 𝗸𝗻𝗼𝘄 𝘁𝗵𝗮𝘁 𝗧𝗲𝗮𝗺𝘀 𝗺𝗲𝘀𝘀𝗮𝗴𝗲? "It doesn't work, here is a screenshot" This weekend I was that colleague: to an AI agent. I had Claude build zoomable mermaid diagrams for our Docusaurus docs for hyground . It nailed the implementation. But the zoomed view rendered completely black. 30 minutes of me describing the problem, taking screenshots and Claude suggesting fixes. Nothing worked. Then I gave it Puppeteer. It took a screenshot of the rendered page and then inspected computed styles: a Docusaurus CSS override killing the background. A 3 lines fix. There's no way the agent could've found it without direct access to the actual outcome. This is what's missing from many agent workflows: agents can write code, but they can't see what it actually does. They're debugging from a description instead of observing reality. To improve outcomes of your agents, give them access to the finished result, so they can autonomously test it.

A few weeks ago I posted about pointing Claude Code at a raw Sega CD disc image of Snatcher and having it reverse engineer the graphics. That was the easy part. Since then I've been using ralph loops, a technique invented by Geoffrey Huntley , to push deeper into the game. Ralph is essentially a bash loop that keeps restarting an AI agent with a persistent task until the job is done. Each iteration gets a fresh context window but reads its progress from disk. I started bottom up. Background graphics were already done, so the next target was animation graphics, then animation timings and positions, then scene data. Every time I hit a wall I realized the same thing: I need more information about how the game actually works. So I'd let Claude disassemble more code, reverse engineer more logic, and then work my way back up. One problem kept coming back. Claude would convince itself that a task was done by claiming the exported data "looks correct" when it was clearly broken. So I built a human feedback loop. The loop would request a review asynchronously through a review tool, essentially making human feedback a tool call. My feedback was treated as ground truth, the same way you'd treat a unit test result. That single change improved output quality dramatically. Progress was actually going well. The loops exported many correct animations and scene graphics from the game. But at some point I realized that working bottom up would always lead to me hitting new walls because only part of the game was understood. So I flipped the approach and worked top down. I set up a Ralph loop that had Claude produce a fully commented and documented disassembly of the Sub-CPU code and every other source file. That became the source of truth for everything that followed. Then I started another loop to re-implement the ASM in TypeScript while loading assets from the game binaries. A port of the actual assembly code into readable, structured TypeScript. First results were rough. So we built a system where Claude can trace the scene VM code through the original ASM and compare it against an actual trace of the TypeScript code. The ASM trace is always assumed correct. Claude would fix the TypeScript bit by bit until the traces matched. Eventually the intro sequence would play but had specific issues with palettes, animations, and scene loading. So I set up agent-browser with a debug system that lets Claude jump to specific frames, inspect state, and step through frame by frame. I gave it the bug descriptions and instructions to track down missing or diverging implementations. It found them one by one. Sometimes it got lost, but that was usually a sign that it needed better debug tooling, more precise bug descriptions, or some steering on how to approach the analysis. The intro is looking almost perfect now. I just discovered that the VM data files contain scene-specific ASM code that needs to be disassembled and re-implemented as well. The loop is already running.