ARCHIVES
The Birth Of (Yet Another) Dark Software Factory
It started with a chat box wired to Claude Code, a Home Assistant API key, and a self-modifying web app I then had to convince to harden itself. A year on, it's a factory that specs, plans, builds, and reviews with the lights mostly off.
devtun: Public HTTPS URLs for Local Docker Containers
I went from ngrok to Tailscale Funnel to Cloudflare Tunnels with Traefik. The last one worked great - but I needed my LLM agents to set it up autonomously. So I wrapped it all in a CLI.
The Human Metaphor Is Breaking Multi-Agent AI Selection
LLMs are overconfident by design. Self-scoring agents inherit that flaw. I tested an orchestrator alternative - fewer tokens, lower cost, better picks.
Why Your LLM Keeps Inventing JSON Keys (And What I Did About It)
I was spending more time debugging JSON parsing failures than building features. After benchmarking 10 LLMs across 4 scenarios, I discovered the real problem wasn't the models - it was my schema design. Here's what I learned.
Stop Manually Tracking Development Ports: I Built an Automated Solution
Tired of manually tracking ports across development projects? I was maintaining a text file like a caveman until LLM agents forced me to automate. Here's devports, the npm package that eliminates port conflicts across git worktrees and lets AI agents manage multi-service development environments without manual intervention.
3 Months of AI-First Development: The Real Results and What I Learned
The honest results from building with AI-optimised architecture using TypeScript, Vite, and Radix UI: what worked, what failed, and the hidden dangers that almost derailed everything.
How AI-First Architecture Made Me 3x Faster: The Design Decisions That Changed Everything
After ditching Next.js for radical separation of concerns, I'm shipping features 3x faster with AI assistance. Here's the architecture that made AI actually useful—and why returning to 'waterfall' practices is the secret.
Understanding OAuth 2.0 Refresh Tokens
A comprehensive guide to OAuth 2.0 refresh tokens using a theme park analogy. Learn how refresh tokens balance security with user experience, and understand token rotation for detecting compromised credentials.
Social Engineering an LLM
LLMs are getting better, they say. And I agree. I'm finding them to be more helpful with coding now than a few years ago. They retain context a little better, drift less, and tend to hallucinate less. But what happens when someone with malicious intent uses the same techniques that make LLMs helpful to manipulate them into crossing ethical boundaries?