Writing

Part three gave one agent a spec. A factory runs many of them on one line, and the idea is fifty years old. What a software factory is, why every era rebuilt one, when it fails, who runs them, and one real factory followed from a whiteboard to production.

Part two ended on a question nobody has solved: how do you know the agent understood what you meant? Spec driven development is the current answer. Where it comes from, who is pushing it, the five approaches worth knowing, and what to do with it today.

Part one ended with five things I want from any agentic harness. This is what each of the five actually costs to build, which part of each one fails quietly, and how to check whether the harness you are using has them.

Google switches off the free Widevine Cloud License Service on 13 April 2027. A second date, two months earlier and unverified, may decide whether your encrypted library survives at all. What to check this week, what the alternatives cost, and which part of the work eats the calendar.

I spent an afternoon designing the wrong tool. The right one was much smaller: there is nowhere to paste a Common Media Client Data payload and read what it says. Along the way I found that version 2 quietly rewrote the format of the keys you already parse.

A model that can write anything still cannot open a file, run a command, or know when to stop pretending it did. The first post in a series on what actually turns a language model into something you can safely put to work.

I grepped my own name in my own codebase and found fourteen hits, all of them in code comments. When I asked the agent that wrote them who they were for, it told me: itself.

The dead time while a coding agent works adds up to a real part of the day. I built wordwright, a terminal typing trainer that teaches grammar and writing in the gap, and this is how its interface is put together.

AWS emailed me a $95 million bill at 3am. I froze my cards and went to sleep. By morning it was 24 cents. The July 2026 AWS billing bug, and what it really taught me.

An experiment. Can an ordinary speech engine catch a single mispronounced sound in Russian and English, and can it hear prosody, the melody and rhythm that carry an accent?

An afternoon experiment: wiring an Amazon Echo to drive Claude Code on my own Mac. The fun was not the voice, it was designing a planning conversation around an LLM that is too slow to answer inside Alexa's eight second window, and then thinking about what a real product would need.

Why letting an AI review its own code is marking its own homework, and how Greenlight enforces a TDD trust boundary between the agents that write tests and the agents that write code.

Give bare-metal services real external IPs with MetalLB as a Layer 2 load balancer for the Raspberry Pi cluster.

Set up an NFS server on a worker node and wire it into the cluster as default storage with PersistentVolume and PersistentVolumeClaim manifests.

Install Kubernetes on the Raspberry Pi cluster: containerd, kubeadm, the Flannel CNI, then initialise the master node and join the workers.

Find your Raspberry Pi nodes on the network with nmap and connect over SSH, both by password and with RSA keys.

Flash Ubuntu 22.04 onto SD cards and prepare each Raspberry Pi node: locate the disk, write the image with dd, and enable SSH.

The first in a series on building a Kubernetes home lab with Raspberry Pi: the prerequisites and the equipment to get started.

How your data hops across untrusted middleman nodes to reach you, who gets to inspect it along the way, and why that quietly shapes your behaviour.

Breaking down what a VPN really is by taking each word in turn: what makes a network, what makes it private, and what virtual actually means.

How middleware can tame the complexity of AWS Lambda functions by isolating business logic from cross cutting concerns, and a look at Lambcycle, a declarative middleware built around a configurable event cycle.