How to use git worktrees to enable multiple AI coding agents working on different features simultaneously - each with their own branch and working directory.

Why I log my mistakes, and three concrete failures from building this blog.

Use a tiny local LLM to intelligently route requests to the right model pool, eliminating wasted API calls

Create a macOS user with sysadminctl. Assume home directory exists. It doesn’t. Unlike Linux useradd -m, macOS requires you to create it yourself.

A model ID that’s ‘close enough’ isn’t close enough. When provider lookups fail silently until they don’t.

How to run AI coding agents in tmux sessions so work continues even if your SSH connection drops - plus recovery commands for when things go wrong.

MCP config that works in Claude Desktop fails in OpenCode. Same protocol, different schema. A tale of subtle incompatibilities.

The button is visible. The selector finds it. But clicks fail silently. The culprit: an invisible overlay you never saw coming.

A pattern for safely executing dangerous operations like firewall rules on remote servers - verify everything before the point of no return.

Reddit’s Responsible Builder Policy blocks OAuth app creation. But the public JSON API doesn’t require OAuth at all - and it might be all you need.