Designing an eval harness for prompt-injection detection: what measuring my defenses actually taught me

AI can help you write code incredibly fast. That part is no longer surprising. The harder question...

There is a scene that is becoming more and more common in software development right now. You have...

Submission for the DEV x Sanity Challenge, Path One. What I Built JudgeStack answers...

I use Excalidraw a lot at work to draw diagrams, plan projects, and visualize how new features will...

Four files, fifteen minutes, and five of those minutes are one step. Plus the four claims about Claude Code setup that get repeated everywhere and are wrong.

The failure is never that the model lies about the transcript. It is that a clerk said something that is not a yes, and the model read it as one.

A step that could not run is easy to report as a step that did not fail. Here is how to tell the difference, and a ten line hook that stops the most common version of it.

A few weeks back I watched an agentic pipeline burn a month of budget in three days. The instinct...

I have a small drawer of Hacktoberfest T-shirts. Four of them, I think. Every October for years I...

Everything your AI agent remembers becomes stale eventually. Here's the engineering behind forgetting: decay curves, core memories, and why we never hard-delete.