Part two ended on the saying half - a reviewer that found eighteen defects and mentioned four. The obvious culprit was the confidence number it files them against, which never varies. So I spent an afternoon and ten cents replacing it with a genuinely calibrated one, and it lost to keeping everything. This is the negative result, why it was cheap, and the one number in it I did not expect.
Every product I start needs the same platform work before the first real feature. I got tired of building it, so I pulled it out into one repository you can copy.
Part one ended with a pipeline that found real defects a prompt never did. So I spent the next stretch of work making the discovery half better, and made the reviewer worse - twice. This is where the constraint went, the ten configurations I ran to find it, and the one prompt rule that was quietly burying the answers.
Last Light has had a PR review skill since almost the beginning, and everyone from Anthropic to Matt Pocock ships something similar. Then I built a golden dataset from real reviews on a real repo and found it was catching one finding in twenty five. This is what I found when I went looking for why, what I replaced it with, and why running the same benchmark twice mattered more than anything I added.
A short primer on the term that has crept into half of my recent posts. What an AI coding agent actually is, why generality is both its feature and its problem, and what a harness adds on top. Meant as the warm-up before the deep dives.
Last week Last Light's coding-agent runtime got swapped twice in four days. Claude Code went out the door, opencode lasted three days, and the thing I ended up shipping is a tiny npm package on top of earendil-works/pi. Here is the story and how it works, step by step.
Writing this from a very long overdue holiday back in far north Queensland in Australia - a tour of where agentic engineering is heading, and what happens to small teams when our job stops being to write the change, and starts being to build the engine that produces it.
Exploring what AI agents mean for individuals, teams, and organisations - from a talk I gave at the Nearform Futureform event in London.
I've spent 20 years trying to answer one question: how do you give users analytics on their own data without building an entire BI platform? Turns out, eventually you just build the platform.
A detailed walkthrough of adding a semantic analytics layer to a Next.js app using Drizzle Cube - from database schema to interactive dashboards.
Building a complete, self-hosted product analytics platform using Cloudflare Pipelines, R2, Apache Iceberg, and DuckDB.
A new chapter, and a time for reflection
A practical guide to getting started with Claude Code, even if you've never written a line of code. Well, almost.
From discovering Superpowers to SVPG with Rocket Skates
How AI-enabled teams can change how we think about product development
A deep dive into implementing rich table editing functionality in TLDraw using TipTap extensions, custom toolbars, and floating controls.
A step-by-step guide to implementing drizzle-cube in a production SaaS application, including security, cubes, and dashboard UI.
Is it really vibe coding if you are actively guiding the tool with firm guard rails?
My latest open source project, an embeddable analytics module based on Drizzle.
Exploring how agentic AI systems can amplify human capabilities and transform the way we work, think, and solve complex problems.
How to fully control a Brink Flair 300 heat recovery unit from Home Assistant using Node-RED, Modbus, and a Raspberry Pi setup.
Three presentations from Microservices Day 2017 covering lessons learned from micro-service architecture, scalable systems, and organisational change.
Presentation video from Micro Services day about how architecture and software development organisations intersect.
A deep dive into debugging a critical memory leak in production Node.js applications at scale, from detection through resolution using heap dumps and understanding V8's string handling behavior.
How I extracted core functionality from Compoxure into two reusable libraries - Parxer for HTML parsing and transformation, and Reliable Get for fault-tolerant service calls.
Presentation video from NodeConf EU on microservice composition patterns using Compoxure and development workflow management with Bosco - tools for building and managing distributed service architectures.
Exploring the challenges of microservice architecture and introducing Bosco, a tool to help development teams manage multiple projects and repositories that inevitably emerge from breaking up monolithic applications.
Exploring how to compose distributed microservices into cohesive web pages using Compoxure, a composition proxy that solves the deceptively simple problem of bringing distributed services back together.