Writing

My kids school was rated Requires Improvement and facing re-inspection. The evidence base was 1,650 files and 918 megabytes. No governor was going to read all of it. So we built a tool that could.

The five-step evolution of how I reach the development environment on the Mac Mini from an iPhone in a playground. The useful insight is that where you build shapes what you build.

Stop hoping that relevant information will fit in the context window. Start manufacturing a task-specific context file before the task begins. The mechanism, the receipts, and the benchmark that gates it shipping.

Long-running tasks fail silently if the session dies before the result is ready. This is the runner I built to make "email me when done" actually mean that. Retry loop, fallback email paths, and a last-ditch file.

Why the regression models that came out of the hackathon got replaced within weeks by three agentic tools. The short version: probability scores without narrative are not what analysts need.

A file I read at the start of every session, twenty-three numbered patterns of how I have broken my own system, and the pre-flight skill that checks proposed work against them. The pattern is the most portable thing on this site.

A two-tier retrieval system (semantic plus keyword), canonical topic files as curated truth, and a nightly consolidation pass that promotes session insights into the canonical tier. Why each piece exists and what fails without it.

What backups, what intentional exclusions, and a sequence that reconstitutes the whole personal AI operating environment in under an hour. The honest version, including the accepted gaps.

A vague ask ("give me a list of prospects that look like X") turned into a working pipeline across three data sources in under sixty minutes. A small build, but the speed is the point.

My seven-year-old had a specific game in mind. Sixty conversational turns later, it was playable on her iPad. A short story about the easiest, most rewarding build I have done.

A one-day hackathon in February 2026, a merchant-risk problem I had been looking at for a year, and a pipeline that scanned 30,000 websites in an afternoon. The origin of everything else on this site.

What happens when one person uses an AI coding assistant as the primary interface to a real physical and operational life, and systematically fixes every failure that occurs along the way.

Every automated helper I built to make my system more reliable ended up making it less reliable. Why, with dated incidents and what I run instead.

The printer-safety architecture I now run, the specific incidents that produced each layer, and why the pattern generalises beyond 3D printing.

Running Claude, Gemini and GPT-5.4 on the same question in parallel, blind Round 0, informed from Round 1 onward. What works, what does not, and the confidence trap I did not predict.

The difference between using AI a lot and being AI-native: infrastructure as the binding constraint, not model capability.