The journey really starts back in 2022. As I wrapped my time with Creandum, I had a realisation: The work I had been doing — interviewing customers, trying to tell founders something they didn’t know — was valuable in marketing; but it was really valuable in product work.
My first product idea came quite quickly: HouseHeld. You buy a house as somewhere to live and relax with your friends and family. With it comes a pile of bricks and pipes and circuits that can literally burn down if you don’t get the chimney swept. HouseHeld would make it easier to look after your home.
I did a ton of customer interviews, went through a billion evolutions of the product. And I started building.
From Figma, to Airtable, to Webflow, to Figma again, and finally to Bubble, which felt INCREDIBLY powerful at the time. I learned so much — but even these relatively simple tools felt like clawing every feature together. Real work.
2023
Before I got much further, a founder I’d worked with a decade earlier asked me to join Beyond Work: a new platform to make working with AI like second nature.
I’d dabbled with LLMs before that, but with cynicism and lethargy bred on the back of years of metaverse and crypto nonsense.
The team at Beyond Work showed me what this tech could do. What it was, how it worked, why it worked, where it didn’t. It was a money-can’t-buy opportunity at the cutting edge.
Much of what the team talked about there is still ahead of what most people are capable of today. How might we do generative UI? What should the experience be like?
Summer 2024
I started to experiment with AI interface ideas in Summer 2024. One idea was a cooking app (Cook King) where you swiped right to get more and more exotic LLM variations on a dish. Or swipe up or down to make it more sophisticated or simpler.
And of course, I quickly returned to HouseHeld…
I put the brief into Claude and asked it to prototype an app and could not believe what I saw. Yes, it was just local storage and technically a Claude Artefact. But it built something real that was way ahead of my Bubble experiments.
2025
My first email about Lovable is from the first week of January 2025. But I think I really dug in at their Hackathon in February. For the AI Tinkerers Hackathon in March, I built the Procrastinorio, an experiment to help you get unblocked on your to do list with Warioware-style exercise — but I built it entirely with Claude Code…
My initial feelings were like wrangling with a creature that was much bigger than me. There was a lot I didn’t totally understand. But I could glimpse a degree of power that was beyond anything I’d used so far.
25th of April, I started a little WhatsApp chat with fellow AI-dventurers. It has been a brilliant space for us to share what we’re learning, celebrate and commiserate through everything going on.
Then in May, I built a whole working app with Claude Code in a month. TradesPurple was the moment I got truly disciplined, starting with a solid scope, truly understanding how to use these tools to get the most of them.
This was before so many of the affordances now. You basically had to tell it every single time “do not write any code yet, just think about how to make this feature and ask me any questions, write a plan and a to do list in a markdown doc and tick things off as you go” etc.
But the power was clear.
Over the summer, it suddenly became easy. I started helping businesses build with AI. I helped the guy I buy wine from in France rebuild his website to escape an ancient web host.
Then in August — I built my first bunch of subagents for writing and a kind of consultancy OS. I started using Claude Code for real work.
If the coding performance had been impressive, the fact I could spin this up and have it write planning documents, emails, etc for me was really unexpected. Kind of shocking.
I spent the next couple of months helping found a startup which has discovered a truly unique way to build the web: full apps using solely HTML + HTTP. And because of all my practice, I could advise them what this tech would mean — AI agents could natively read and use any website or web app. (PS THIS IS NUTS, expect to see more very soon.)
Meanwhile, an old friend now working for a $2bn tech co asked if I could take some of what I learned with my orb_OS and AI content experiments and help them out.
What’s incredible about working at this time is: we agreed to build something like the subagents + output-style I had done before. Then, the day before I was due to start building, Anthropic released Claude Skills.
And from the Monday to the Friday, we had built a Claude Skill that anyone in the team could use by just dragging a zip into their desktop app. And write SIXTEEN THOUSAND WORDS in 30 minutes, to a standard their in house writer said was 99% there. (It now does more documents in half the time, as well as details like media drivers, UTM codes etc, and has no usage limits.)
When the rest of the marketing team heard about this, it led to more projects. So now I’m bringing AI to one of the biggest British companies in the world. And don’t tell anyone, but I’d probably do it for free, because it’s absolutely bloody brilliant.
In the last months of the year, I started booting up Claude Code inside my Obsidian folder — a notes tool that stores all its files in markdown, a simple text format that AI basically loves. I got a plugin that downloads all my meeting notes as markdown files there too. And a to do list app which syncs with a text file there.
With a few Claude Skills, this has become a stupendous work machine. We talk about my business goals and plan together. I sit at my desk on Monday and ask it what I should focus on this week, what I should do today.
The next step was I asked it to digest everything in my notes, with a particular focus on meeting transcripts, and build a “work graph” of who I work with, our relationships, what we’re working on, what companies are involved etc.
This got really fun. When we talked about work, I could then ask it to update any of those notes. It really shows the benefit of running LLMs in a constantly growing repo of knowledge instead of chat by chat. Apparently, Anthropic is building this into CoWork next — I wouldn’t be surprised at all.
It spawned my latest project: Knot. Knot is a notepad that remembers and understands your work. When you first start it up, it creates that work graph I described above, from your meeting notes. Then, every morning it boots up a live, editable notepad of what you need to keep in mind that day.
Not a daily briefing — not a static thing. A notepad, like any other text editor that you can and should interact with and change and tweak. It’s an active experience, with two way sync back to your work graph. And after every meeting it gets smarter and smarter.
This is probably what I’m most excited by as we move into 2026. LLM power where the chat is implicit.
2026
The biggest leap forward I’ve seen since working in LLM tech in early 2024 is thinking loops. This is what enriches a basic prompt and still gets you good answers. This is what allows tool use loops that make agentic examples like Claude Code possible.
But it’s also a background task. It happens implicitly without the user, behind the scenes.
I think this shows us the way forward. When do we wake up every morning to a whole night of interesting work or analysis or possibilities from these tools? Not prompted, but proactive.
Spoilers: today. It’s already here. My Claude Skill can do the weekly review, the monthly planning + forecast without me asking.
And so, we might finally get to the real questions about how you manage all this power. I don’t believe normal people will Claude Code their own tools and make their own software. That is a recurring pipe dream from engineers that never comes true.
Instead, someone is going to build the interface for this. And I’d be stunned if it doesn’t look very like Beyond Work’s experiments back in those early days.
Someone is out there working on this right now — probably starting without the baggage of Anthropic or Open AI, who (excitingly) have no unfair advantages to discover this.
Someone is doing the work of a lifetime. And maybe of all our lifetimes.
I just hope I can continue to contribute and help others discover what’s possible.