Blog
Teaching AI to See
Part 4 of a series on Claude API features for builders So far in this series we’ve given Claude the ability to think deeply, take actions, and speak in exact formats. But everything up to this point has been text in, text out. In this post we’re crossing into a different territory: giving Claude eyes. Vision and PDF support are the features that let Claude understand images and documents, not just the text in them, but the charts, diagrams, layouts, photographs, and visual information that make up so much of how the real world communicates. If you’ve ever wished you could just throw a screenshot, an invoice, a floorplan, or a research paper at an AI and have it understand what it’s looking at, this is how you do it.
March 26, 2026
Making AI Speak Your Language
Part 3 of a series on Claude API features for builders In Part 1 we gave Claude space to think. In Part 2 we gave it the ability to act. Both are impressive, but there’s a fundamental problem lurking beneath both capabilities that anyone who’s tried to build a real system on top of an LLM has encountered… the output can be a little unpredictable. (which is part of the point of LLMs right!)
March 24, 2026
When AI Learns to Use Tools
Part 2 of a series on Claude API features for builders In Part 1 I covered Extended Thinking, how giving Claude space to reason before answering dramatically improves the quality of its responses on hard problems. That was about Claude learning to think. This post is about Claude learning to act. Most AI products people use today are stuck in the advice phase, it’s question and response, chat chat. They can analyse, summarise, draft, and suggest, but when it comes to actually doing something, a human has to take the output, leave the chat window, and go execute it somewhere else. Copy the SQL into a terminal. Paste the email into Outlook. Manually look up the data the AI said it needed.
March 23, 2026
What Happens When You Let AI Think Before It Speaks?
Part 1 of a series on Claude API features for builders We’re all inundated with AI-related content these days, but for me I seem to see more AI=bad than anything else online. It’s all jailbreak-this, benchmark-fail-that, incoming apocalypse, and risk risk risk! But some of it will be good. So I’m planning to write some AI=good content, and hopefully give some of you good reasons to explore bringing AI technologies into what you’re building. I’ve been working with the Anthropic Claude APIs a lot lately, so I’m going to focus on some of its features that I think are genuinely useful, interesting, and worth understanding properly.
March 16, 2026
We Need to Build
Give me a 100 million dollars, 20 smart and well-paid people, access to GPUs and cloud resources, and twelve months. I’d give it a 50% chance of being a worthwhile investment, but it could be the Snowy Hydro of the 21st century for Australia. I’ve watched the discourse about AI in Australia for a while now, the hand-wringing about being “left behind,” the earnest policy papers, the endless conferences. And I keep arriving at the same conclusion: this isn’t an AI problem. It’s a risk problem. We’ve lost the appetite for doing big things for the benefit of the nation.
February 25, 2026
On Finding Your Tenets
I’ve been reading Dan Koe lately. If you haven’t come across him, he writes about building a life around ideas, writing, creating, thinking for a living. One concept of his caught me off guard: the idea that articulate people aren’t pulling insights out of thin air. They have maybe 8-10 “greatest hits” that they’ve refined through repetition, ideas they can connect to almost any topic because they’ve thought through them hundreds of times.
December 16, 2025
Securing Firebase Cloud Functions with Workload Identity Federation
If you’re building a mobile app with Firebase and want to protect Cloud Functions that call paid APIs - AI services, databases, or external integrations - you need to understand a critical security gap: Firebase Auth validation happens inside your function, after it’s already running and billing you. This means every request to your function is billable, even if the token is invalid. Malicious actors can spam your endpoints with garbage tokens and drain your budget. Your Firebase Auth validation will reject them, sure - but not before GCP bills you for each function invocation.
November 27, 2025
On Building vs Joining
I’ve been going down the startup path over the last 12 months, a very different life to the decades I spent in big tech companies. Someone recently asked me which was “better”, joining an established company or building your own thing. It made me laugh a little, not because it’s silly, but because I’ve spent the better part of 25 years convinced I knew the answer, only to completely reverse my position… twice.
November 20, 2025
On Legacy
On my last day at AWS, after 8 and a half years, I posted one final message to my team. It was that meme from The Usual Suspects, you know the one, with the caption “and like that, he’s gone.” Then I removed myself from the Slack channel and shut down my laptop. I’d left companies before. Twice from Microsoft, and both times I’d felt that surge of panic, the “oh shit, what have I just done?” moment. But this time, closing that laptop after nearly nine years, I felt… relief. I was ready. I was tired. And more importantly, I was moving on to build something that mattered to me in a different way.
November 15, 2025
On Change
“Everything changes all the time.” I heard this phrase so many times at AWS that it became almost like a mantra. I can’t remember which manager said it first, or even if it was always the same person saying it, but I heard it over and over. In a company moving at AWS’s pace, where new services launched every week and teams reorganized every quarter, it wasn’t profound wisdom so much as observed reality. The thing about hearing something repeatedly in that environment is that eventually you stop noticing you’re saying it yourself.
October 29, 2025