Seedance 2.5 prompts that get liked: labeled blocks and a handheld camera, measured on 582 posts
We matched 582 Seedance 2.5 prompts shared on X with their like counts and compared the most-liked quarter with the least-liked, within each posting month. Liked prompts split into labeled blocks (61% against 48%) and ask for a handheld camera (49% against 34%). Time ranges, 30 seconds and words like cinematic are in nearly every prompt, so they do not set the liked ones apart.
Summary
If you write a Seedance 2.5 prompt, split it into labeled blocks and put a handheld camera in it. Those two habits separate the prompts people liked most from the ones they liked least.
The reason to look at likes this way is simple: almost every Seedance 2.5 prompt already does the basics. We matched 582 prompts that creators shared on X with their like counts, ranked each against posts from the same month, and compared the top quarter with the bottom quarter:
- Labeled blocks. 61% of the most-liked prompts break into blocks such as
[STYLE + CAMERA + ATMOSPHERE]orCHARACTERS:, against 48% of the least-liked. One plain paragraph is the reverse: 21% against 32%. - A handheld camera. 49% of the most-liked ask for handheld, against 34%. Outside vlogs, where nearly every prompt is handheld anyway, it is 40% against 23%.
- Longer. The median most-liked prompt is 2,788 characters in 14 lines; the least-liked, 2,069 in 10.
And three things that do not separate them, because both groups do them: time ranges like 0-5s (55% against 59%), a 30-second plan (the median in both) and words like cinematic (67% against 66%).
Every prompt below links to its video on our Seedance 2.5 page.
Our picks
- Give the prompt labeled blocks. Start with one for the look and the camera, then the characters, then the action by time. The medieval walk of shame opens with
[STYLE + CAMERA + ATMOSPHERE]and then[CHARACTERS]. The pattern held when we kept one prompt per author: 63% of the top half against 41% of the bottom half. - Ask for handheld, even outside a vlog. One phrase is enough. The roller coaster wig says
subtle handheld camera energyin its first line and then times each beat. - Keep the time ranges and the 30 seconds. They do not make a prompt stand out, but nearly everyone writes them, and ByteDance’s own examples set the pacing with ranges. Treat them as the floor. The most-liked post in the set, the bedsheet ghost prank with 19,446 likes, is a timed shot list.
- Spend the words on what happens, not on style labels. Cinematic, photorealistic and 4K appear just as often in the least-liked quarter. The extra length in liked prompts goes into characters, action and camera.
- Judge a prompt against its own genre. Commercials and product spots make up 21 of the least-liked 150 and 6 of the most-liked 150. A product prompt with modest likes may be doing fine for what it is.
Look at your last Seedance prompt: is it one paragraph, and does it say how the camera is held? Those are the two lines to change first.
How we tested
- The prompts. Every Seedance 2.5 entry from X in the open collection awesome-image-video-prompts at commit 590b532: 591 posts by 221 accounts, posted from July 31, the launch day, to November 2.
- Likes. We read each post’s current like count from X’s public embed endpoint on November 7, 2026, one request per second. 582 posts returned a count; 9 were deleted or unavailable.
- Fair ranking. Likes keep growing after a post goes up, so we ranked each prompt only against posts from the same month. The top quarter is 150 prompts with a median of 644 likes; the bottom quarter is 150 with a median of 120.
- Checks. We repeated the comparison three ways: one prompt per author (each author’s best, so prolific accounts count once), English text-to-video prompts only, and prompts outside the vlog category. We call a gap real only when it held in all of them and passed a two-proportion test at about the 5% level.
- Features. For each prompt we counted characters and lines and tested patterns: structure (time ranges, numbered shots, labeled blocks or plain prose), the planned length, 22 camera terms, six style words, audio, spoken lines, negative instructions, an aspect ratio and references. The exact patterns are in the script.
Results
What sets the most-liked apart
| Top quarter | Bottom quarter | One per author, top half / bottom half | English text-to-video, top / bottom | |
|---|---|---|---|---|
| Labeled blocks | 61% | 48% | 63% / 41% | 62% / 38% |
| One plain paragraph | 21% | 32% | 22% / 33% | 18% / 44% |
| Handheld camera | 49% | 34% | 45% / 31% | 52% / 32% |
| Median length | 2,788 characters | 2,069 | 2,752 / 1,977 | 2,724 / 1,589 |
| Median lines | 14 | 10 | 15.5 / 8 | 13 / 7 |
What both groups do
| Top quarter | Bottom quarter | |
|---|---|---|
| Time ranges like 0-5s | 55% | 59% |
| Planned length, when timed | 30 seconds | 30 seconds |
| States a length | 77% | 75% |
| Cinematic | 67% | 66% |
| Photorealistic or realistic | 76% | 71% |
| 4K or 8K | 23% | 21% |
| Aspect ratio | 39% | 44% |
| References | 21% | 27% |
Leaning, not settled. Liked prompts mention sound a little more often (58% against 49%), say what to avoid more often (79% against 71%) and zoom more often (18% against 11%). None of these passed our test, so read them as hints.
By genre
| Category | Top quarter | Bottom quarter |
|---|---|---|
| Action | 33 | 33 |
| Vlog | 32 | 26 |
| Cinematic | 28 | 18 |
| Sci-fi and fantasy | 20 | 11 |
| Commercial | 6 | 21 |
| Music and dance | 3 | 9 |
Is handheld just a vlog effect? We checked. Inside vlogs, 84% of the most-liked and 88% of the least-liked are handheld, so the camera does not separate them there. Outside vlogs, it does: 40% against 23%. Our guess, which the data cannot prove, is the look: a slightly unsteady camera reads as filmed rather than generated.
What to ignore
- Style words. Cinematic, photorealistic, 4K, IMAX and anime are about as common in the least-liked prompts. They cost nothing, but they will not lift a prompt.
- The aspect ratio. Liked prompts name one slightly less often. Pick the shape your platform needs.
- Very long prompts as a goal. Length comes with the blocks and the extra detail; the most-liked prompts are longer because they say more, not because of a length target.
Limitations
- Likes are not quality. A post’s likes depend on the account’s audience, the video and the timing as much as on the prompt. We had no follower counts, so the one-per-author check is our best guard.
- The collection is already selective. 392 of the posts came through a list of popular Seedance prompts, so the bottom quarter is the least-liked of a liked set, with a median of 120 likes.
- We did not generate videos. This says how liked prompts are written, not that the same prompt written another way would get fewer likes.
- The features are text patterns, written mostly for English. A camera move described in other words is missed.
Reproduce it
The script is seedance-2-5-prompts-likes-2026-11.py. Clone the public collection, check out commit 590b532 and run it; it needs Python and PyYAML and fetches like counts from X’s public embed endpoint at one request per second, so a full run takes about ten minutes. It writes seedance-2-5-prompts-likes-2026-11.json: one row per prompt with its likes, monthly percentile, features and a link to the source file, plus every group summary in this post. Like counts will have grown since November 7.