The idea arrived in 2019. Build something that preserves your memories so future generations can actually know who you were. I took it to a couple of developers, we looked at what it would need, and we agreed it could not be built. They were right. Then COVID happened and it went on the shelf.
It came back in 2025. I had been using AI seriously for about two and a half years and the technology had changed underneath the idea. Claude was what made me realise it was now possible.
The concept never changed: a place to keep everything — videos, photos, documents — that learns who you are and can recreate you. Your children and their children could talk to you. Or, if you fancy, ask Steve Jobs a business question and get a real conversation back.
Two products, deliberately. Legacy is public — anyone can look you up and ask you something. Generations is private, for family and close friends, and holds the things you would never say to the world.
The part I feel most strongly about is the digital estate. A will handles the house and the money. It does nothing about your Facebook, your LinkedIn, your documents in the cloud. In most cases nobody can get in, and the platforms keep profiting from you after you are gone. Our answer is a digital executor: someone you nominate who can actually manage your digital life when you cannot.
I started the specification work around Christmas 2025, using Agile, which was a mistake. I had AI generate about 500 user stories with acceptance criteria, then wrote detailed specs from them. What I should have done first was design the interface. Instead I spent weeks arguing about fundamentals that should have been settled on day one.
The platform decision alone changed five times. VR first — the technology was not ready. Then mobile. Then voice-only mobile. Then I remembered people like keyboards. Then it hit me that a native app hands Apple or Google up to thirty percent of everything, which is unworkable, so: web app. Then responsive, for desktop and tablet. Every pivot made the previous specification obsolete.
The turning point was discovering what Claude could do with design. That genuinely changed things. The output was excellent, and because I know how this works I could direct it precisely — workflows, layout logic, interaction patterns. We iterated until it produced something I can only call a masterpiece.
Then I made the next mistake: I wired up all the workflows before making the site responsive. That cost me 850 individual changes. By February I had upgraded my plan. By late March I was running Claude almost continuously.
On the back end I picked a big cloud provider, because that is where my background is. I described what I needed, Claude wrote the scripts, I pasted them in and watched the environment appear. At the peak I had two terminals going at once and eleven agents running — an orchestrator, developers, testers, performance testing, the lot.
Then the expensive lesson. I pushed a container through the pipeline without realising the upload had switched on a service. It ran. It billed. The month closed at a $2,000 bill. After a good deal of persistence I got most of it back, and I am still chasing the rest.
I moved the whole thing to open source shortly afterwards. If that cloud could provide those services, I could build them, and I had Claude. So off I went.
The most interesting part of all this has been watching how I use AI change. I have worked through the model range and learned that the biggest model is not always the right one. I mix them now depending on the job.
The hardest thing remains making everything work properly on every screen size. Multiple layers of checking, including real browser tests, and it is still not quite at one hundred percent. Close. Not there. It has taken an enormous amount of time.
