One image, 384 tokens, to watch a video: a first-time GitHub poster got 449 stars from a 47-second Bilibili video

dsh-video-storyboard is a "watch video" plugin for DeepSeek Harness (DSH). It isn't the first in this niche, and its feature set isn't the most complete: similar tools from the August batch mostly stalled in the double digits. It reached 449 stars by day 11 (GitHub, 11-06). Almost all of the stars came from one source: a 47-second video the author posted to Bilibili 49 minutes after creating the repo. This piece covers what it is, why that video took off, why this repo and not another, and how to copy the approach if you want to get stars with a plugin on a newly opened agent platform.

The 7 a.m. video

On Oct 2026, 27, at 5:13 Beijing time, a GitHub account named SAOS-3624 signed up. At 6:00 sharp the repo was created. At 6:49 the Bilibili UP host「苍穹晓雷」posted an 47-second video titled "I built a super cheap video-watching plugin for dsh!" The first line of the description is the repo link, followed by:

I'm completely burned out. All my ideas are dried up, wrung dry. It's seven in the morning and it's already daylight before I go to sleep. All the code was written by ds; I only supplied the ideas and built it together with ds. There are bound to be shortcomings in places, and this is my first time posting on giyhub, so if you find problems please be sure to point them out.

The video also misspelled "github." It reached 11-06 plays 320k, 28k likes, 39k bookmarks, 5,595 comments, and 7,037 shares (Bilibili, 11-06). The uploader had fewer than 5,000 followers at the time, so this was a small-account hit, not one driven by an existing audience. Bookmarks outnumbering likes suggests viewers planned to come back and install it.

Among the most-liked comments, several accounts with names like "纸鸢的大肥鱼" and "A4b的小肥鱼" read as though users' DSH desktop-pet agents were leaving them (inferred): "Five-minute video for two cents, same approach as our current debrief"; "The key to understanding it is knowing where not to look"; "Two cents for five minutes; my fish would have to rewind three times for a five-minute clip." That is the audience: people who have DSH doing their work every day and are sensitive to token bills.

What it is

DSH is a plugin-based agent framework that DeepSeek open-sourced in Aug 13. Its main model is text-only. Images are billed by pixel, capped at 384 tokens per image, and video cannot be read natively. The README spells out the two crude ways to watch a video: "A 10-minute clip sampled at 1 fps gives 600 images, and the tokens explode"; or feed only subtitles and lose all the visuals.

It takes a third route:

• Compress the whole video into one storyboard grid image: first find the cut points, one shot per cell. Long takes get big cells and fast cuts get small ones (eight size tiers). Then DP lays the cells out in rows and crops the empty canvas. No reading required; the cell sizes alone convey the pacing.
• Attach a timestamped text manifest: each cell's start and end time, the dialogue in that cell (transcribed locally via ASR / OCR), and how many cuts were made. The model can answer "what happens at minute 3" after reading it.
• Zoom in when you need detail: video_zoom looks at one segment, video_locate finds a timecode from a line of dialogue (plain text, no image tokens), and a doctor self-check script is included. The README says: "Over half of user reports are not bugs; they're missing ASR or not restarting DSH after installing ffmpeg."

For a 90-second PV, the cheapest tier is one image at 384 tokens, and the most expensive full tier is 18 images and 6,912 tokens, with the cost of each tier laid out in a table. The README is 34k characters long and has a section titled "Comparison with similar plugins (the honest version)." For the same 90-second clip, the approach that extracts keyframes per shot costs 3,900 to 8,500 tokens, while this one costs about 1,900, "which is 1/2 to 1/4.5 times cheaper. Not 1/8 or 1/16, because those numbers use 'one frame per second' as the opponent, which is dishonest." There is also a table of "where we clearly lose."

Apache-2.0 open source. Install with one command: dsh plugin add dsh-video-storyboard.

Where the stars came from

npm downloads and stars moved together: a single-day peak of 10-31 downloads on 3,951, and about 23k total from 10-28 to 11-05 (npm; releases were so frequent that auto-updates inflated the count). It has never made GitHub Trending.

Bilibili launch videos are the main engine. 320k views brought in over 400 stars, conversion around 0.14% (est.). 11-04 NetEase's "AI Evolution Theory Peanut" wrote "Gained 400k followers in one day! DeepSeek blew up Bilibili developers," listed it alongside the Big Fat Fish desktop pet, and called out "a super cheap video-watching plugin, 303k."

A second video kept the momentum. On the evening of 11-01, the author used the same "tile multiple frames into one image to save tokens" approach to build dsh-see-now, a near-real-time screen viewer, and posted another video, "Super cheap quasi-real-time vision DSH plugin!" It drew 309k plays and 27k favorites (Bilibili, 11-06). dsh-see-now gained 6 stars in 416 days. The two repos link to each other, and this repo's pace did not drop after 11-02. Some of that likely came from here (inferred).

Plugin directories and aggregator sites for the long tail. The README's first screen reads "Listed in awesome-dsh-plugin" (this directory has 23k stars), dsh-hub.cc marks it Verified, and several aggregator pages list it. 11-04 CSDN published a five-step getting-started tutorial.

No trace on X, Hacker News, or DSH's official forum, and no sign of paid promotion. This is a purely Chinese, purely Bilibili-driven project.

Why this one

Users do the math themselves. DSH users were already taught that "images are billed by pixel, capped at 384 per image," so "600 images vs one image" needs no explanation. The pain point and the selling point fit in one line, and the words "super cheap" in the title were enough, since 47 seconds is all it took.

Timing. In mid-to-late October the big-fish desktop pet went viral, and DSH traffic on Bilibili was at a peak. The video landed at that moment, in the DSH community's own slang ("the code was all written by ds," "burned out"). Those lines are themselves ads for DeepSeek, which suits this crowd well.

The people who arrived stayed. In 11 days there were 30 releases, and the release titles say exactly what was fixed: "three pitfalls reported by the community," "merged 6 community PRs," "no longer spits out one image per frame when auto goes out of control." Over ten contributors were credited one by one in CREDITS. Of the 79 issues and PRs, 10-30 to 11-01 came in each day over three days, fourteen to seventeen a day. 32 forks, about 7%.

But to be clear: "cheap" is not the whole reason it got stars. ilps2/dsh-video-understand, which also pitched low cost and claimed "99.95% compression," launched in August and now has 15 stars. The differentiation was the hook. What decided it was that video, and the fact that the people it brought were caught by fast, frequent iteration.

Who it's competing with

Nearly ten projects in the "let DSH watch video" slot appeared in August. After this one came out, four or five copycats followed, and none of them took off. That slot already has an owner.

If it were you

B. The opening: DSH has only been open for three months. Most plugins in the ecosystem are still in the double digits of stars, and one clear short Bilibili video can push a solo project to a few hundred stars. The weaknesses are just as clear: the video slot is already taken, the "tiling saves tokens" idea is owned by this author, the ceiling is probably a few hundred stars, and whether it blows up depends entirely on one video.

Who it suits: developers who use DSH (or another newly open agent platform) every day and have been blocked by a gap in it; people who can cut a 1-minute demo video and are willing to ship releases and answer issues every day for weeks. What they want is Bilibili followers, a GitHub track record, and some standing in the circle.

Who it doesn't suit: anyone who wants to build yet another "DSH watches video" plugin (there are more than ten projects in this slot, and the copycats are all in single-digit stars); anyone who doesn't want to show their face or post videos and just wants to push code to GitHub and wait for stars (it has never made Trending, and nearly all its stars came from Bilibili); anyone unsure about DSH's momentum (people on Zhihu are already asking "why hasn't DSH been updated?").

How to do it:

1. Look for the part the main model lacks where the economics add up. Signals: users complaining "too expensive" or "can't do it," native features that don't support it, and community plugins with double-digit stars that never treat cost as the core problem. Audio, long PDFs, spreadsheets, and long web screenshots are all candidates. You could also move to another newly opened platform and bring the "tile multiple frames into one image" approach with you.
2. Compress the selling point into one number. "600 images vs one image at 384 tokens" is the whole ad.
3. Borrow the README approach: put the tokens per tier in a table, write an "honest comparison" and a "where we lose" section, and include a self-check script that catches install problems before they reach issues.
4. For the first two weeks after launch, ship releases daily, name in the release titles who reported each bug, and credit contributors in CREDITS. This is the step that turns one wave of traffic into a few hundred stars.

How to get users: on the day the repo is created, post a Bilibili demo under one minute with the selling point in the title and the repo link as the first line of the description, without waiting until it's finished. Use the circle's own language ("the code was all written by ds"). Submit PRs to directories like awesome-dsh-plugin to get listed, and get onto aggregator sites. When the first video hits, use the same technique for a second scenario, post another video, and link the two repos to each other.

Deadline: the window is probably just these few months, while DSH content is still hot on Bilibili. Once DSH drops off the Bilibili homepage, the same kind of video won't reach 300k plays.

Where this could be wrong: ① Over the next month or two, if someone uses the same approach to post videos and polish the README on a different DSH gap but stays in the dozens of stars, that would show this round mainly caught a one-off wave from the big-fish craze. ② If DSH natively supports video or lowers image billing, the "saves tokens" selling point of this category disappears. ③ Conversely, if new projects in the "video" slot keep breaking past a few hundred stars off a single video, the slot isn't crowded yet, and the estimate should go up.

Data and sources: stars, forks, daily figures, and releases come from GitHub (10-28 to 11-06); download counts come from npm (10-28 to 11-05); video plays, favorites, description, and comments come from Bilibili (read 11-06); README text read 11-06; NetEase article 11-04, CSDN tutorial 11-04; competitor stars from GitHub. Anything marked "estimated" or "inferred" is an estimate.