Burning Tokens Made My Eyes Water
Last updated
Table of contents
How the site started
Almost as soon as ChatGPT landed in my hands, I had an idea to make a website, thennow.nz. The first commit is from May 2023. I did not understand anything about React development, and I still don't, so I built the whole site by copy-pasting with ChatGPT. I managed to take a few photos and drive around a bit, but then I got very busy and forgot the site existed.
I came back to it now and then, took a few more photos and played with artificial intelligence a little to measure how much easier it had become. It had become much easier. For example, I no longer had to copy anything into a chat and back. Everything worked on the computer from the terminal.
The live site on a phone before the T3 runs, built with help from ChatGPT and Codex in 2024 and 2025.
Recently something pushed me to come back to it again. A good friend of mine reminded me about it when she suggested I buy frames with old photos of Auckland, because she remembers better than I do what I like. She said: you like old Auckland, you can hang the frames at your place. I didn't buy the frames, because they turned out to be cheap crap from Kmart. But I am grateful to my friend for reminding me about my website.
Two tricks this week
Since I am an autist obsessed with artificial intelligence, x.com, Telegram and the rest of the internet are, for me, one continuous stream of news and content about AI. Lately I keep seeing skills and methods for improving UX and UI. This week alone I came across two. I tried them today on my website.
One of them was, I think, the Dennison skill. I asked ChatGPT to set up two sessions in Claude Code. One session worked in vanilla Claude Code, and the second used Dennison’s skills.
Vanilla Claude Code.
Claude Code with Dennison’s skill pack.
I did not notice any difference in design improvement when I used the skill. I think I would have got a slightly better result if I had typed in by hand what exactly I wanted from the redesign, and then the skill would not be needed anyway. So either I could not apply this skill, or it is not powerful enough, or I am using it wrong. I am sure Dennison is a respected person. He has his own products, he is rich and successful, but somehow I just could not appreciate and apply his skill properly.
Theo
The second skill I found by accident, literally today. More precisely, it found me: it turned up in my x.com feed. It comes from Theo, a developer, investor and who knows what else, as he calls himself. He has a development environment, T3 Code. I saw glowing reviews of it from respected people on x.com and installed it too. And I really liked the editor, from installation and onboarding to actual use. I think I will use it for a while. The quality of the product hooked me.
Theo’s pinned post.
So when I saw a new pinned post from Theo in my feed, saying he knows how to save a whole lot of developers by sharply improving UX and UI at the cost of burning a lot of tokens, I got interested. It was perfect timing. I had just been reminded about my hobby project, I had spare tokens left before my subscription reset, and I had a website that obviously needed a dramatic UX and UI improvement.
My spare Codex tokens that evening.
And my spare Claude tokens.
For scale, Theo’s own token bill: about $75.6K. His caption about Claude Opus 5.5, two thirds of that bill: “it has its quirks but it’s the greatest workhorse of all time”.
What I asked GPT to do
Again, I did it very simply. I opened the same chat with GPT that I had been using before. I told it: launch T3 and run an experiment there. Redo the homepage the way Theo recommends in his post on x.com. Theo’s method is also published as a skill.
I did not burn as many tokens as Theo might recommend. My model chose for itself how much to spend on iterations. I think I had 4 or 5 experiments for the homepage and for the stories, while in screenshots and reviews other developers wrote that they did 10 experiments, if I remember correctly, 10 iterations. And maybe they used more powerful models. I used only Claude Opus 5.5 and GPT-6.1 Sol. I don’t even know at what power. The agent decided that for me too, naturally.
What it looked like on my phone that night: a new review session for every round.
Each reviewer scored the design blind.
One of those other developers: Kaelig on X.
The result
After the T3 runs, live on thennow.nz now: the homepage, then a story.
I’m ready to pray to tstack for this alone: it figured out by itself that inside a story it should show the original photograph and the modern shot separately, each on its own.
Why I am writing this
I am writing this post because it blew my mind. I almost cried from an overflow of feelings. The jump in UX and UI quality was just impossible. For the effort spent versus the result, I have never done anything like it in my life. It feels like magic. I did not suffer, I did not write, I did not slave away. I just sent a screenshot of Theo’s post on x.com and got an explosive jump in UX and UI quality, without even explaining what exactly I want. I just said: make it pretty and send me the results.
I can’t even imagine what will happen if I do really deep research, understand what I actually want from the site and give the agents a direction. I’m afraid I’ll just lose sleep for a week. I don’t think I’ll survive it, because I have already lost sleep over what just happened. I’m dictating this post at half past midnight because I can’t sleep. There is one more reason I can’t sleep, but I can’t tell you about it. It’s my little secret. And it has nothing to do with artificial intelligence, thank God.
Anyway, a chaotic post, but I want it chaotic, so it comes from the heart. And I beg you: if you have pet projects at work or at home, or any interface with UX and UI, install T3 at any cost and spend 10 iterations improving some part of the UX and UI that is dear to you. I bet you will either shed a tear or have some other strong emotions, and in short, you will be very surprised by the result.
See you next time. I’ll try to go to sleep. And I’ll spend the last half hour editing this post for my website and for a LinkedIn post.