How a Karpathy-style LLM knowledge base became the place where I think, plan, and draft — with agents reading from it instead of starting from zero.

I stopped letting one model do the whole loop. Now Opus plans, Sonnet implements, Codex reviews, Opus fixes and closes — and Hermes tells me what happened.

I connected Sentry to Claude through MCP. Now instead of staring at dashboards myself, I ask Claude to go look at what’s broken, find performance problems, and come back with a plan. The interesti...

A failed Slack migration, a dead daemon, and two weeks of specs written without a laptop — how CodeButler v2 became a multi-agent team.

Three CodeButler daemons, one phone, and a bug fix deployed from a beach towel without opening a laptop.

How CodeButler manages Claude sessions — continuation markers, input waiting, and a compact watchdog that keeps context bounded.

Adding a local TUI input and bidirectional image support to CodeButler — including a raw mode rabbit hole.

How a simple idea — text my phone, have Claude Code do the work — took five architecture rewrites to get right.

What happens when you ask the same creative question to three different AI models simultaneously? I ran an experiment using OpenClaw: I asked my AI assistant to spawn three parallel sessions—one f...

I used to think the hard part of working with agents was prompting. Get the words right, get the output right. It seemed like a skill you could grind: learn the tricks, master the phrasing, unlock ...