One 82-second film, six platforms, six different scores

On 10 October I published a film. It is 82 seconds long and it draws an astrolabe. An orange spark reads sixty lines of TikZ and drafts the instrument: the rim and its scales first, then the plate worked out for Oxford's latitude, then the rete. When the pen stops, the rete is cut out of its construction lines and time starts to run. It turns nearly six times and settles on sunset over Oxford at 18:22 on 10 October, the day it went out. The music was generated with YuE2 from the film's own structure; the sound effects are code, a pencil for every stroke.

When I reviewed it I gave it 95 out of 100. For the first time the picture, the sound and the words all agreed with each other. I thought of it as a finished piece of work, and all it needed was people to see it.

That day it went to six places. Bilibili got the Chinese and English versions as two parts of one upload. Xiaohongshu got the vertical film and, separately, a set of nine stills. Douyin got the vertical film, YouTube the full film plus a Short, and X a seventeen-second highlight. An agent filled every upload form in the browser, and I clicked the last button each time.

Two days later the platforms had scored it very differently.

Xiaohongshu did best. By the evening the video note had 380 plays, 26 likes and 15 saves. Its sister film, a San Yuan feng shui compass, had a harder time there: I deleted it twice because the first two posts never entered the feed at all, then posted it a third time under a plainer title, "An ancient compass". That one had 1,556 plays and 68 saves within a day.

Douyin did worst. Fewer than 2% of viewers watched the minute and a bit to the end.

Bilibili sat in between and hurt the most. The series, Drafting Sheets, has never taken off there. By 10 October its first four episodes had 461, 140, 194 and 52 plays, each lower than the last. The astrolabe had 112 by the evening of its first day. I didn't quite accept that, so I bought one round of Bilibili's paid promotion. Plays went up by a thousand. Likes and saves barely moved, and a few people turned up in the comments to criticise it. I hid the video, and most of the money came back.

That sent me into Bilibili's creator dashboard, which compares every upload with similar uploads. By its own measures the films are fine. Average watch-through sits around 40%, against a median of 34.6% for videos under three minutes. The engagement rate is between 10% and 20%; the median for comparable videos is about 1.5%. People who see these films finish them and save them. What's missing is people seeing them: the first batch of viewers the recommender hands out is tiny.

It took me a while to read the numbers properly.

The first episode of my other series, The Order Beneath, is about complex analysis. On Bilibili it has 7,471 plays and 1,171 saves, my best result so far. Its average watch-through is only 23%, below the median for its length. But it is eighteen minutes long, so 23% is still a little over two minutes. A Drafting Sheets film, even watched to the end, holds someone for half a minute to a minute. My working theory is that Bilibili cares about how much time a video keeps people in total, and a short film can't add up to much however well it's watched. That theory rests on a dozen uploads of my own; the algorithm won't tell me if I'm right.

Other work fits the same pattern. A music video I posted at the end of September, two minutes forty, has 280 plays on Bilibili and 4,959 on Xiaohongshu.

Douyin's 2% is a different problem. People there are flicking through one video after another, and if the first two seconds don't hold them, they're gone. The astrolabe opens on a sheet slowly lighting up and drawing itself stroke by stroke; the rete doesn't start turning for nearly fifty seconds. That pace barely stands a chance there.

The YouTube story is a little embarrassing. A dozen videos had almost no views for days, and a friend searching for their titles from another account couldn't find them. Every one of them had been marked "made for kids", which on YouTube switches off comments and keeps a video out of personalised recommendations. That setting was mine to choose: the agent leaves it for me on purpose, and I had been choosing wrong all along. Once it was changed, the channel was open for business.

At the beginning of October I wrote down what I wanted my content to be. On 5 October the line was: make works, treat every piece as an asset, and see the day it's published as where it starts. What I wanted was a name that someone searching a keyword a few years from now couldn't avoid.

This week made that line more concrete for me. Whether something is an asset depends on whether it's still there in a few years and whether people can still find it. The astrolabe has its own page, where the film is drawn live in the browser with its full listing beside it, and that page will stay. The posts on each platform sink within two or three days. I still collect the platforms' numbers every day, but they're only worth collecting if they change how the next film gets made.

These are the rules that came out of the numbers over the past ten days.

On 7 October, Bilibili's dashboard showed three episodes of The Order Beneath losing 33% to 36% of viewers in the first three seconds, against a median of 24.7%. Since then every film opens on its best frame and moves within a second, and a script measures the first three seconds of every export. A film that fails doesn't ship.

On 10 October the astrolabe was due to get a fifty-second vertical cut. The music is one continuous piece, and cutting the film into spans cut the music mid-phrase too; it sounded broken. I dropped that version. From now on the vertical and the horizontal versions are both the full film, plus one vertical highlight that plays a single moment.

The same day I noticed that the Xiaohongshu image post carried exactly the same text as the video post. Image posts now get their own text, one line per still, in order.

No more paid promotion. It buys plays, not viewers who want to keep watching, and it puts the film in front of people who never wanted it.

From here Drafting Sheets will probably go out in two forms. The short, wordless films will live mainly on Xiaohongshu, with YouTube and X alongside. On Bilibili I plan to post only long versions: the film first, then the same renderer pausing and explaining how an astrolabe is used, why a plate only works at one latitude, and how the code turns circles in the sky into lines on brass. I haven't decided about Douyin.

The long versions are expensive in tokens. I'm on a single 20x account. Two short films a day it can carry; two long ones a day it can't come close to. Posting five episodes of The Order Beneath back to back used up a reset. So things will slow down a little: one or two long episodes a week.

The astrolabe is still hidden on Bilibili. Its page is on lookbook.catforest.dev, and whoever opens it gets the whole thing drawn again in their own browser, from the first stroke.