Writing
On technology, leadership, AI, and the messy reality of building things that last.
38 posts
Black October sent me to the Three Lines model, the risk framework auditors use. It says the first line of defence isn't the CISO. It's the people who run the service. I run a web platform, so that's me, and AI just made my response time the whole game.
More Writing
AI
In February, Accenture made AI use a condition of promotion. In October, it began centrally funding only $29 a month of Claude per person for many employees. The cost cut isn't the story. The yardstick flipped, and the people who followed it are holding the risk.
AI
An end-of-life Drupal 7 site became a static Astro site with the same look, URLs and content in two days, about eight hours of my attention and less than $100 in GitHub Copilot token costs. The agent started from the open-source unDrupal skills, could commit nothing until a person accepted it, and migrated what visitors saw, not what the database said.
AI
I put three plain-English browser tests on this site, judged by Cloudflare's Clef decision model through Sedum. The most useful results were the first failures: a section hidden until you scroll and a click the runner refused to guess, each reported as a number below a threshold.
AI
How I'm wiring Asana, Copilot's cloud agent, GitHub Agentic Workflows, Deployments and Playwright into one delivery pipeline where every request, agent and test result lands on a single pull request. Plus why decision models like Jev and Clef belong in the QA gate.
AI
Chat models write sentences for people. Decision models like TypeSafe's Jev return typed, calibrated answers for code. The difference isn't speed or price. It's who reads the output, and I learned that by pointing Jev at the wrong reader.
AI
GitHub Spec Kit hit 1.0 last month. Most of what it does is tell a frontier model how to work, and frontier models stopped needing that. What survives is narrower: intent a person reads before the code exists, and a test that makes the goal checkable.
AI
The most useful thing I read this year about resistance to AI is a 2,400-year-old claim that nobody chooses harm as harm. It explains the team that won't adopt the tool, and it describes a misaligned model exactly.
Physics
In Alcubierre's warp solution the ship never moves — space moves around it. I built a 3D model of the metric to see the mechanism, and the slider that refuses to move turned out to be the more useful object: a way to sort what is stopping a technology into knowledge, energy, matter, or time.
AI
Three hours of Microsoft agent training, then an evening in the Scout docs — where I found my own setup staring back at me. The agent loop has commoditized. What Microsoft is actually selling is the boundary it runs inside.
AI
About half of what AI generates for me is usable — and the accept rate isn't measuring what I thought it was. A systems read on why verification cost, not model quality, decides where augmentation actually pays off.
AI
Cloudflare's CFO says non-human traffic could outnumber human traffic 1,000 to 1 within five years. That doesn't create a new problem for enterprise web presence — it makes an old one impossible to ignore. Here's what I'm doing about it.
AI
HBR says AI is rewriting the economics of outsourcing. I've been living the micro version of that argument for the past year — here's what it looks like from inside a web team.
AI
How a Claude Project turned nine nights, three lodges, and a folder of confirmation PDFs into a reconciled itinerary — plus a homepage update pushed from the flight home.
AI
Moving a domain from .com to .ai is a trivial technical change — and a marker of everything that changed between hand-cut XHTML templates and directing AI every day.
AI
AI was the unsanctioned helper every company quietly encouraged. Now finance found the invoice. What the tokenmaxxing reckoning says about the bigger bet.
CMS
Adding a visual web editor to a static Astro site — three wrong turns, a TypeScript surprise, and a pivot before I even got started.
AI Development
A practical comparison of AI coding tools in 2026 — Claude Code, Cursor, Copilot, Codex, and cloud agents — with an interactive decision tree to find your fit.
Writing
Why I write what I do — and just as deliberately, what I don't. Selective disclosure, vulnerability that earns its keep, and an AI editor keeping the line.
AI
How AI moved from chat assistant to force multiplier in my work and life — and the first-principles insight that finally unlocked CI/CD after three years of failed attempts.
AI
What For All Mankind, the Luddites, and a kid named Alex tell us about AI — and why this wave of automation may be the first one workers can turn on their own behalf.
AI
AI has made the possible infinite. That makes saying no the most important skill you have. Opportunity cost, token budgets, and the strategy of choosing what not to do.
AI
After months building a personal agent in OpenClaw, I gradually realized Claude Code was doing everything I wanted — and more. The framework that drew me in lost its edge when the tool it was built around became the whole stack.
AI
Using Claude Code and an open-source SEO skill to audit quevin.com — and discovering that Google had never even crawled it. Here's what we found and what we fixed.
Leadership
Three ideas I keep returning to about how organizations actually work — and the gap between what leadership says and what culture does.
Health
After discovering elevated Lp(a) and borderline Stage 1 hypertension in my labs, I started using Claude as a health co-pilot. Now I'm texting it my meals to track sodium intake.
AI
Combining Obsidian as a human reading layer with a Claude Code agent as an automated organization engine — a documentation system that lives in git, organizes itself, and compounds in value over time.
AI
A decade-old idea about delegation and getting unstuck just got a lot more powerful. AI doesn't just expand what you can get done — it makes the practice of delegation infinite.
AI
I now have two Claude-powered bots on Telegram. They do very different things. Here's how I set them up and why both exist.
AI
A five-minute Telegram conversation turned into a working GitHub Action that auto-posts new blog entries to Bluesky — with images — and backfilled all eight existing posts.
AI
One month after building a personal AI agent with OpenClaw: what changed, what works, and what it's like when your own agent drafts a post about itself.
AI
AI is burning people out — not because it doesn't work, but because it works just well enough to remove the natural limits that kept us healthy. Neuroscience and an Aldous Huxley quote explain why.
Drupal
How we used Claude Code with Opus 4.6 and an AI-augmented 20/60/20 workflow to build a CKEditor 5 plugin connecting Aprimo's DAM to Drupal, delivering responsive images with srcset and lazy loading.
AI
From skeptical experimenter to daily collaborator — how Claude became embedded in my workflow as a technical lead, and what that actually looks like in practice.
AI
What it actually takes to run your own always-on AI assistant — the setup, the surprises, and what's still unresolved.
travel
A 12-day family trip to Brisbane, Australia—featuring koalas, sea turtles, New Year's Eve fireworks, and how I used Claude AI to plan and document the journey.
AI
Learning to treat AI coding assistants as amplifiers, not replaceholders, through a cost-saving Drupal 7 to Astro.js migration project
Astro
Why I chose to migrate from React to Astro, and how markdown-based content management changes the game for personal websites.