BLOG
Writing about AI agents, architecture, and the craft of engineering
A Good SKILL.md Is the Cheapest Reliability Upgrade You'll Make
SkillSpector tells you whether a Claude Code skill is safe. Nothing told you whether it was any good, or how much its prose was costing you. So I built the quality layer and benchmarked it: across Haiku, Sonnet, and Opus, a strict schema and one worked example moved structured-output correctness from 0% to 100%.
Claude Code Best Practices for Vibe Coders: Ship More, Burn Fewer Tokens
Vibe coding with Claude Code is fast until the token bill isn't. After a year of daily use, I found the habits that fix both sides: the explore-plan-code workflow, path-scoped rules that trim CLAUDE.md by 60%, deterministic hooks that enforce guardrails at zero token cost, and the context tricks that keep long sessions from derailing.
AWS Summit Hamburg 2026: The Year Agentic AI Went from Hype to Production
From automotive supply chain AI to a self-healing enterprise HR platform: real companies, real results, and the honest questions still left open.
Modern React Performance Without the Overhead
INP replaced FID in 2024. The React Compiler handles memoization. RSC payload size is the bottleneck most teams miss. Here's how to find and fix all three.
turbovec: Local RAG Without the 60 GB Tax
float32 embeddings eat RAM fast. turbovec cuts a 10M-doc corpus from ~60 GB to ~7.5 GB at 4-bit, with no training pass and unbiased similarity scores.