196 days, 47.9 billion tokens, one person's project matrix

I keep a token monitor open on my machine. When I took a screenshot on 13 October, the total read 47,940,460,187, about 47.9 billion. Next to it: 196 active days.

A week earlier, on 5 October, I had written "40 billion in half a year" into my content strategy. Eight days added nearly eight billion more. This month a day over a billion tokens is normal; the busiest day reached two billion. After that peak the trend line runs downhill and the last few days sit near the floor. A friend's birthday and a three-day weekend came along just as the quota ran dry, so I took a few days off. On 1 September I crossed 880 million in a single day for the first time, and I thought it worth putting on this site's timeline.

Under the total the monitor shows a dollar figure, $40,222.05, its estimate from usage. What I actually pay for is three subscriptions, one 20x plan each for Claude, Codex and Cursor, and most weeks I use up every refreshed quota. I'm setting up a second Claude 20x account.

By model, Opus 5.5 accounts for 15.9 billion, a third of the total. Opus 5 is 13.4 billion, then Opus 4.8, Fable 5, and on the Codex side gpt-6-astra with 2.9 billion.

Where it goes is easiest to show with one day, 10 October. In the small hours I signed off an 82-second film that draws an astrolabe, and every version of it was exported in one go: horizontal, vertical, a highlight, nine stills, covers in four ratios. Two repositories each shipped a release that day, one of them to give the content ledger a new kind of video version. From the morning into the evening, the film went out on six platforms. An ordinary day looks much the same: a bug in my day job being fixed in one session, older content being reworked in another, the next film in production alongside. Several things move at once and I switch between them.

I can work like that because I spent a long time building the ground first.

In June I made no products at all. I spent the whole month on a Next.js starter, a base layer of UI components, an API layer the projects share, and yBlocks, the component library I later built pages from. From the outside nothing happened that month. Without it, the dozen-odd projects that followed could not have run in parallel.

The rules were written one by one in the same way. Data tables, forms, boards and settings pages each have a fixed way of being built, with working examples in the component library. Colour, icons, loading states and keyboard shortcuts each have their own rule. Before work starts on any project, the AI reads these files, so I don't have to explain again what a button should look like.

Beyond the rules there are procedures. Every repository keeps a feature ledger, one entry per finished capability, and yPulse collects all of them every night. Every repository has a release card that says where its version number lives, which commands gate a release, where it ships and how to check it afterwards. I say "release", and the dozen steps after that follow the card. Publishing a film works the same way: the job sheet comes from the ledger, the AI fills each platform's form in the browser and stops at the publish button, I make the last click, and the link goes back into the ledger.

Chained together, these are the core of the matrix. What I have in mind the whole time is a one-person company. The component library, the rules, the ledgers and the pipelines cover what a company would split across several departments, and the AI carries them out. My part is to set the rules, make the calls, and stand at a few gates. Problems found in the data take the same route: when Bilibili's dashboard showed too many viewers leaving in the first three seconds, that finding became a script that measures the opening of every exported film, and a film that fails doesn't ship.

Right now the whole thing is limited by tokens.

Two short films a day the quota can carry. Two long ones a day, like the twenty-minute mathematics films of The Order Beneath, one 20x account can't come close to. Posting five episodes of that series back to back used up a reset. To stretch the budget I changed the division of labour last week: the main session only makes decisions, and the mechanical work of running gates, exporting films and filling forms goes to a cheaper model, with the logs and screenshots kept out of the main session. That saves a good deal, but long films are still limited to one or two a week until the second account is running.

Since Opus 5.5 arrived I have felt two jolts. The first was in video production. The second came this weekend, when I saw a set of animated SVG figures someone had made with Claude: name a concept in one line, and Claude invents a machine that demonstrates it, built from dozens to hundreds of parts, with every readout coming from the same model that moves them.

The most complex of them, a turbofan engine, took Claude fourteen hours by its author's count.

I'm still working out how to get that second Claude account.