Updates from Magnus Hultberg
I ported my Cloudflare opinionated AI coding template from Claude Code to Mistral Vibe CLI. The methodology came across cleanly. Switching to a different harness imposed a lot of practical change.
Your file system is chaos, and no AI model is going to tidy years of it into a single source of truth. You don't need it to. Pick one corner with decent content, ship something useful, and widen from there.
Product Managers are posting their GitHub commit counts with real pride lately. I share that sense of pride, yet I don't write a line of the code. Some of it is maybe pure vanity. But some of it I believe is making me a sharper Product Manager.
I spent a weekend building a thousand-year-old Viking game in the browser, and ended up using a graphics trick from 1993 to make it look good. Turns out Claude knows about Doom. Because of course it does.
Two-thirds of the field notes in my Victorian creature generator started with "Upon". Adding "vary your openers" to the prompt helped for an entry or two, then the model parked on a different shape. Some problems with LLM output cannot be fixed by rewriting the prompt.
The first time someone runs Claude Code in a project folder, the AI doesn't ask what they want to build. It asks who they are. An update on my opinionated project template for shipping web apps with Claude Code, four projects in.
🔧 Most of us assemble a stack, get it working, and never question it again. The same instinct that makes you challenge a feature requirement applies to your tech choices too, and with agentic coding acting on the answer has never been cheaper.
I built an AI reading triage tool to get on top of my Readwise Reader queue. It worked, briefly. Then Jevons' Paradox paid me a visit. 🧙♂️
🧠 I spent a weekend building a memory system for Claude Code. It had a serious case of amnesia. Turns out I'd built a highly capable (and seriously over engineered) motion sensor for a room with no motion.
There's a maker 🛠️ revolution happening around AI tools that reminds me of the early web. Here's how I'm using it, and where I draw the line.
👍 After a year of AI-assisted building, the interesting question isn't if the model can write the code. It's whether you've built the conditions where something good can emerge. Project Ansible taught me that the right constraints aren't a cage. They're what makes you brave enough to experiment.
Working with Claude Code doesn't feel like programming. It feels like sketching. Fluid. Creative. Playful. I'm molding clay, not writing code. 📡 Project Ansible (AI summaries for Readwise) is my most ambitious build yet, and it felt like painting.
After more than a year of experimenting with AI assisted coding 💪 the interesting part for me is no longer “can the model do this?” but “what’s the structure and process that actually makes this work?” This project template is where I have ended up, and that now kick starts my projects.
Seven or eight real apps built with Claude Code, zero “actual” coding from me. A new white paper just described my workflow so precisely it felt like reading my own brain - so here’s how I actually do it.
Just grokked: CLAUDE.md files are sent with EVERY request. My rich 700+ line memory for 'do this task' seemed suddenly wasteful. Refactored to act as library index. 📉 Result: 730 → 442 lines (67% reduction). 30 min discussion, 2 min execution.
What if the best way to get great code from AI is to make it politely argue with itself? I tried the "antagonist agent" approach from a Claude Code event — and it genuinely feels like having a team. 🤖⚔️🤖
Given the right setup and guardrails coding agents become more and more productive, as shown in a recent article from OpenAI. What I find striking is how well their findings for maintaining context and high quality output align with how a high functioning team would work.
China's betting on AI while the West argues culture wars. Are we heading for Banks' utopian Culture — or something darker? Two reads that made me feel like we're standing at a civilisational fork in the road.
The "AI Swiss Army knife" hype isn't empowering young talent. It's a cost-cutting memo dressed up as a compliment. Generalists have always been the glue of great teams. Let's not confuse accessibility of tools with depth of craft.