AI Engineer, IT Consultant. Building AI agents that run in production.
I scored the AGENTS.md files of the 16 biggest AI agent repos with a deterministic engine. ~1.46M combined stars, zero A grades, and the repo that popularized the convention ranks 13th of 16.
Writing an agent loop is one line. Engineering what it carries between wakeups — idempotency, pacing, convergence, a cost ceiling — is the job.
OWASP LLM Top 10 risks live in your instruction files. Runtime evals don't lint them. Here's what static security analysis catches — with a concrete finding.
A CLAUDE.md I wrote was merged into modelcontextprotocol/servers. My own scorer returned 59.2/100. Where it was right, where it was structurally unfair, and what the gap to 80 would cost.
Earlier today I published a self-audit that named a blindspot in schliff. Hours later, the fix is up as a PR. The same CLAUDE.md now scores 61.0 instead of 59.2.
I built an AI matching system, ran it into Annex III, and walked out the other side. A solo builder's walk-through of EU AI Act compliance — what changed, what I filed, what nobody warned me about.
Your CLAUDE.md is an agent harness. Skills are tools, hooks are guardrails, memory is persistence. How to stop configuring and start designing.
Karpathy's LLM Wiki solves a problem older than computers: knowledge maintenance. Why every system from Memex to Obsidian failed — and what's different now.
A code review council built on Karpathy's LLM Council — six advisors across Claude Opus and Codex, cross-examined and synthesized into one verdict.
I stopped comparing Claude Code and Codex last Tuesday. That's when Codex started reviewing Claude's code — from inside Claude Code.
A curated guide to the Claude Code skills that actually save time — from automated debugging to spec-driven workflows.
MCP has 97 million monthly downloads and 10,000 servers. But for 80% of what developers actually do, curl and jq is faster, cheaper, and more reliable.
Claude Code hooks are shell commands that fire on events inside the agent loop. A working guide to the schema, the full event list, and the four traps people keep hitting.
How Claude Code skills became the new dotfiles — a personal developer identity layer that defines not just your tools, but how you think.
A practical look at Model Context Protocol servers — which ones matter, how to configure them, and what they enable.
Every non-trivial article on this blog starts as a markdown spec. A practitioner's case for spec-first work with AI assistants — with the artifacts named.