AI Wrote the Code. Do You Understand What You’re Shipping?
Decide what you expect before reading an AI agent’s solution. Then check the code, the assumptions, and what the tests actually prove.
Notes on production work and personal projects, with implementation details, tradeoffs and limits.
Subscribe via RSSDecide what you expect before reading an AI agent’s solution. Then check the code, the assumptions, and what the tests actually prove.
Three design directions, one editorial layout, and a photo I decided to remove. What redesigning this site taught me about giving AI useful feedback.
An agent can write code, add tests, and still solve the wrong problem. Faster implementation leaves plenty of product decisions to make.
Everyone's talking about which model is best. The real differentiator is memory - what the agent remembers, how it remembers, and what happens when it forgets.
I spent weeks wondering why some Claude Code sessions felt fast and cheap while others burned through my usage. Turns out, prompt caching is the invisible infrastructure that makes the whole thing viable.
I built a group chat where the AI doesn't wait for slash commands. It sits in the conversation, decides when to speak, and pulls out decisions and tasks automatically.
Something changed in late 2025. Here's what actually happened, what it means for engineers and companies, and my honest take on where this is heading.
I've been running OpenClaw for weeks and kept hitting token limits without understanding why. So I sat down and dug into every layer: sessions, memory, browser automation, costs. Here's what I found.
Everyone's talking about AI agents. Most predictions are wrong. Here's what I think will actually happen, based on patterns I'm seeing now.
I've been coding with AI assistants for months now. Here's what actually changed, what's overhyped, and why I'm more optimistic about engineering than ever.
The future of AI isn't one super-intelligent assistant. It's specialized agents working together. Here's what I've learned running a personal multi-agent system.
Live encoding updates at VdoCipher, using SSE and Kinesis during the team’s AWS-to-GCP migration.
Notes on the search work at VdoCipher: response times from 2–3 seconds to about 200 ms, and the checks that matter around a search index.
Frontend build time fell by about 90%. What I worked on, and what a build-tool migration needs to preserve.