Writing
Everything I have published here, newest first. Filter by type or by tag.
32 pieces
One message became nine features, three pull requests and 557 passing tests in production. The building took a day. Deciding what was worth building, and whether it was any good, took the same as it always has.
I rewrote Split Race from a 30-second top-down toy into a 3D racer with real tyre physics. Every bug that mattered showed up in a headless simulation, long before anything was drawn.
A short dictionary of the things people say at work instead of the things they mean. It is funny until you notice what the vagueness is actually doing, which is moving the risk onto you.
I spent two months more productive than I have ever been. It moved nothing on paper — no raise, no progression. Here is the lesson I came back with, and why I stopped confusing output with leverage.
The model is a commodity you rent by the token. The part that decides whether any of it works is everything around it.
A structured workflow for identifying at-risk customers before they leave - combining usage data, support signals, payment behaviour, and engagement patterns into a repeatable detection and response system.
Leadership and a job title are two different things. One is granted. The other is practiced. And once you notice the gap, you cannot unsee it.
Anthropic shipped its most capable model on June 9 and pulled it on June 12. I used it the entire window. Here's what changed in my day-to-day, what didn't, and the real reason it's gone.
When you're chasing a performance problem, the high-level stats will always send you in the wrong direction. The answer is never in the summary. It is always in the most granular thing.
Knowing how to use AI tools is table stakes. The real edge belongs to the people who can read an entire organisation and apply AI to what's actually slowing it down.
A workflow for detecting, naming, and stopping scope expansion before it kills delivery. I just put this together - time to test it out.
Most enterprise data infrastructure was designed for dashboards, not decisions. Agents expose that gap fast.
AI writes most of the code now. I still build by hand, and not out of stubbornness. Here is what the craft loop keeps giving me.
The multi-agent setup was working. Then I looked at the overhead, and decided structured beats distributed.
A look at the matching engine behind the Ask widget on the CV page - how it normalises input, scores intents, and decides what to say.
Clear ownership is not a nice-to-have on improvement projects. It is the single most important thing. Without it, the project will stall, conflict, or quietly die.
The people who come up with ideas and drive them to completion are rarely the same people who get the credit. That gap is real, it is common, and there are practical ways to close it.
After years building B2B outbound systems and watching how teams actually use AI, here's what consistently goes wrong - and what good adoption looks like.
A case for technical ownership: why reaching for Next.js and a database often beats adding another subscription.
A structured decision-making process that forces choices through progressive layers of thinking - clarifying the problem, defining success, exploring options, comparing trade-offs, stress-testing assumptions, and producing a decision with a clear execution path.
I let a team of AI agents manage my portfolio site. Here's what that actually looks like — and what I've learned so far.
The full decision process - identifying what to automate, choosing the right tool, mapping the logic, testing it, and handing it off so it doesn't break the moment you stop watching.
The question isn't which is better. It's which is right for what you're actually trying to solve.
Most AI pilots look great in the demo and die in production. Here's why - and what to do differently.
A comprehensive workflow for building top-down, bottom-up, and cross-departmental communication that actually works - maximising transparency, protecting employee voice, and making it structurally difficult for leadership to ignore the people doing the work.
A complex, production-grade workflow for designing, deploying, monitoring, and deprecating AI agents - covering decision gates, dependency resolution, orchestration, and feedback loops.
Why unifying your tools isn't a technical problem - it's a strategic one. And why most organisations get it backwards.
Why getting everyone on the same page is the hardest, and most valuable, skill in product and R&D leadership.
A repeatable workflow for translating complex R&D work into clear, decision-ready documents that anyone can act on.
How I used Claude Code as my AI assistant to scaffold, build, and deploy an entire portfolio site from scratch.
First post. Setting up a dev blog powered by Next.js, MDX, and Vercel. Here's why and how.
Practical tips for building a dark theme that doesn't feel like a cave. Terminal aesthetics, readable text, and accent colors.