Alex Lewis

A completeness check that had to be safe over time, three state machines that each leaked a fail-open, and the pure function that finally fixed it.

LLM-as-judge is the default eval now, but a judge is just another unreliable model. Here is how to check whether its verdicts correlate with real outcomes, and catch the leniency bias most judges have.

A year ago I kept a 6,000-line migration coherent with a hand-maintained progress file. Most of that scaffolding now ships inside the agent CLIs. Here is what the harness absorbed, and the part it still cannot.

The biggest productivity gap is the work around the code: context gathering, note-taking, and status updates. I built 50 AI workflow skills to make those 15-minute tasks actually happen every day.

I ran /insights in Claude Code and walked through what it found - friction patterns I hadn't noticed, CLAUDE.md rules I should have written weeks ago, and what changed after I applied the suggestions.

This is the third post in our series on probabilistic prompt pipelines. In our first post, we explored why static prompts become bottlenecks. The second…

This is the second post in our series on probabilistic prompt pipelines. In the first post, we explored why static prompts become bottlenecks and saw a…

This is the first post in a four-part series on building production-ready probabilistic prompt pipelines. By the end of this post, you'll understand why…