Lessons learned, problems and solutions
What it really looks like to develop a project from scratch with an AI as a permanent pair — architecture decisions, real limits, patterns that work. Lessons learned from sharebox, a self-hosted file sharing tool in pure PHP.
How to measure AI prompt quality, build an automated evaluation harness, and iterate to a quality plateau. 5 versions, 150 roasts, concrete lessons on what AI does systematically and its limits.
Your git log and memory files are the most honest source for creating custom Claude Code skills. How to audit them, identify recurring patterns, and write skills that auto-trigger at the right moment.
Bug fix, new feature, refactoring, code review: each dev task has its own prompt template. I tested and scored dozens of formulations — here's what actually works with current AI models.
Same machine, same Node.js, two approaches: a cron to publish on a fixed schedule, a daemon to react in 30 seconds. Cron, systemd timer, daemon — when to use which and why.
Everything Claude knows about how you work lives in ~/.claude/ — and it doesn't sync. How to rebuild your global CLAUDE.md, reference files, statusline and plugins in minutes on a fresh machine.
Claude Code's auto-memory keeps across sessions what evolves too fast for CLAUDE.md: preferences, active decisions, gotchas. The 4 types, filtering criteria, and an audit prompt to keep it clean.
.claudeignore, domain-split CLAUDE.md, aggressive /clear between tasks, targeted questions — concrete levers to cut token usage without losing quality.
How I built a bot that scrapes LinkedIn posts, detects new comments, and replies via Claude CLI in under 30 seconds.
REST for public APIs, gRPC for internal services. Server streaming for a real-time price feed, Protobuf for typed contracts, and the 4 streaming modes explained.