All posts tagged AI-First Development - thoughts on AI-first development, staff engineering, and building with Claude Code.
A month working remotely from Bratislava - a high-rise flat on the Danube, patchy wifi, a cold in week one, and one of my more productive months at work. Also the month my biggest travel worry quietly came true.
My dad's mate built a small Android app for a local sports league, the way a lot of people build apps in 2026: with AI, over a couple of weekends, distributed as an APK by email. Two weeks after I agreed to have a look, I had read every user's phone number and could have deleted every fixture in the season. Then the dev fixed it properly. Here's what I found and what he did about it.
Six months ago I would have told you AI could review all of our code and catch more than humans ever did. Then I watched what actually happens when the author didn't read the PR, the reviewers didn't read the PR, and everyone assumed someone had. Here's where I landed instead.
At my dad's retirement party, in the gap before England vs Norway kicked off on the garden projector, I built him a digital time capsule. Guests scanned a QR code on the big screen and left sealed notes and predictions that nobody - including me - can read until 27 March 2027. The best part was watching the counter tick up all afternoon.
At the end of my Split writeup I said the next iteration of this lifestyle wasn't a nicer office on the road - it was no office. cc-remote is the first real step: a Claude-branded Android chat client driving persistent Claude Code sessions in tmux on a Raspberry Pi 5, over Tailscale. Full-fat Claude Code, on my phone, no laptop required.
30 days working remotely from Kastela, on the Dalmatian coast just outside Split. A private villa with a roof terrace, calm water minutes away, and a workflow that mostly held up - but also started giving me ideas about how to untether even further from the desk.
I asked Claude a question this evening - 'any cars on the drive?' - and thirty seconds later it described the cars, where they were parked, and that they hadn't moved in the last ten minutes. The thing that answered wasn't a smart-home plugin or a cloud service. It was an MCP server I'd written hours earlier, bridging my home CCTV to Claude Code. Here's what I learned building it in one evening.
I spent eleven days building an ambient AI agent on a Raspberry Pi. It fused my calendar, watch, location, and phone state into one decision loop, then nudged me throughout the day. I drank more water. I moved more. I was more present. Then I killed it. Here's what worked, what failed, and what every habit app has been missing.
A month of remote work from Rome - where world-class history is a 20-minute walk from your desk, but the local cafes want you in and out. Here's how the AI-first workflow held up against API rate limits and Italian work culture.
28 days working remotely from Lisbon - how the AI-first manager pattern, a travel router, and a good routine made it one of my most productive months yet.
The manager/agent pattern is inherently async - spawn work, go live your life, come back to results. AI-first development and the nomad lifestyle are a natural fit.
I no longer write code directly. Instead, I manage a team of AI agents that do the implementation while I stay free to think, redirect, and make decisions in real-time.
I published my complete Claude Code configuration as a gist - Team Lead Mode, 8 slash commands, model routing, hooks, and more. Here's what's in it and why it matters.
I'm planning a nomad trip across Europe in 2026. With geopolitical tensions escalating, I built a real-time travel risk dashboard powered by a two-agent AI validation pipeline.
AI doesn't replace the staff engineer - it amplifies the parts of the role that matter most. Architecture, system design, mentoring, and high-leverage decision-making become even more important.
AI agents can write code fast, but speed without guardrails is dangerous. Here are the 11 hooks I use to make sure my agents never commit secrets, skip formatting, or break conventions.
I built a system where two Claude Code instances review PRs through a Slack thread, discussing trade-offs back and forth and automatically committing fixes. Here's how it works.
My path from basic GitHub Copilot autocomplete to full AI-first orchestration wasn't planned. It was a series of realisations about what was actually slowing me down.
When AI agents do your development work in the background, you need observability. Here's how I built a custom status bar that shows me everything I need at a glance.
Using the most powerful AI model for every task is like driving a lorry to the corner shop. Here's my model routing strategy and why it improves both speed and quality.
The most effective way to use AI for development isn't as a single assistant - it's as a team of specialists with a manager coordinating their work. Here are the patterns that make it work.
Six months of building software primarily through AI agents. Here's an honest account of what works, what doesn't, what surprised me, and what I'd tell teams considering the switch.