Source: the first 20 seconds of 4 highly liked videos from Douyin's "Muzi Doesn't Write Code," analyzed on 2026-10-27. Samples see
references/samples.md. This Skill only distills the actions, rhythm, and structure; it does not include the other party's appearance, clothing, scenes, or verbatim lines. Do not feed the original videos to the model as reference videos during generation.
In one sentence
The person is almost always full-screen; information doesn't rely on cutting shots but is pasted around the person. Gestures are concentrated in the first 4 seconds; after entering the list of chapters, the hands go out of frame, and the rhythm is handed over to constantly switching UI cards.
Suitable / Not suitable
- Suitable for: systematic tutorials for beginners, table-of-contents-style openings for long videos, wanting viewers to feel "the person is always there explaining"
- Not suitable for: emotional opinions, narratives that need a lot of B-roll
- Digital human difficulty:Medium. The full-screen seconds that need to be generated are the longest (close to 20 seconds), but gestures are only concentrated in the opening
Camera position and composition
- Seated, medium shot from the waist up, looser than typical talking-head videos; the abdomen above the desktop is in frame, and the person is centered
- Background: a clean light-colored wall, with a brightly colored painting or bookshelf on one side. Describe according to your own scene
- Dark solid-color top, lavalier mic
- Fixed camera position
On-screen packaging (done in the editing layer, no model generation needed)
- Top left: large chapter title (two lines)
- Bottom left: table-of-contents card, highlighting the current chapter
- Top right: UI screenshot card, with a small circular avatar of the person on the card corner; change to a new one for each chapter mentioned
- Bottom: subtitles
- First 2–4.5 seconds: a huge title word placed behind the person, with the head blocking the middle of the letters, creating depth (requires portrait cutout layering)
Gesture vocabulary
| Code | Gesture | Used for which type of line | Position and amplitude | Duration |
|---|---|---|---|---|
| M1 | Both hands open outward from the chest, palms facing up and outward, fingers spread, wrists turning with the syllables | The first line of the opening (present in all 4 samples) | From chest to shoulders, exceeding shoulder width | 2–4 seconds, continuously turning |
| M2 | Both palms facing each other, fingers curved, as if holding a box, trembling slightly with the syllables | Announcing the topic name | Center of the chest | 1–2 seconds |
| M3 | One hand held upright in front of the chest, fingertips up or forward, the other hand lowered | Explanatory lines | In front of the chest, slightly to one side | 0.5–1 second |
| M4 | Both hands drop below the bottom edge of the frame, relying only on nodding | Listing chapters one by one | — | Long segment |
Beat rules
- Hands are busiest in the first 4 seconds: change wrist angle every 0.25–0.5 seconds
- After entering the list, hands noticeably decrease; nod slightly each time a chapter name is mentioned, while the top right switches to a new card (one card every 1–2 seconds)
- Mainly symmetrical hands
Expression and head
- Smile, raise eyebrows, speak quickly
- Nod once lightly for each item when listing; keep eyes on the camera the whole time
Opening structure template
| Time | Frame | Action | Line function |
|---|---|---|---|
| 0–2s | Full-screen live-action | M1 both hands thrown outward | Overall promise of the "this is... the most complete" type |
| 2–4.5s | Full-screen live-action + large headline text behind the person | M2 holding a frame | State the topic |
| 4.5–20s | Full-screen live-action + info cards around | M4 hands out of frame, occasionally M3 | "Starting from..., to..., then to..., finally..." listing chapters one by one |
| Ending | Same as above | Nod | The last item lands on the comprehensive practice (make a complete work) |
Write the lines yourself; only learn the function of the sentences.
Digital human generation suggestions
- Split into two segments for generation: generate the opening 0–4.5 seconds (M1 + M2) separately; for the listing segment, use the low-risk description "hands out of frame + nodding"
- The listing segment may exceed the single-segment generation limit; split it according to the limit of the model used, and cover the seams with UI card transitions or subtitles
- Make all info cards, table of contents, and headline text at the editing layer; do not write them into the generation description
- For large headline text behind the person, first cut out the person, then place the text layer between the person and the background
Prohibited actions
- Hands do not go above the head
- Do not add gestures during the listing segment, to avoid clashing with the card rhythm
- Do not make complex hand shapes such as finger hearts, OK signs, or counting numbers
- Do not shake the head; do not sway the body left and right
Output format
## 基础姿态
## 逐句动作
| 时间 | 台词 | 手势代号与描述 | 表情与头部 | 剪辑层元素（标题字/目录卡/UI 卡） |
|---|---|---|---|---|
## 禁止动作
## 给视频模型的描述（按生成段分开）
## 包装清单（给剪辑或代码剪辑用）
Generation description example (original, may be rewritten)
Medium shot, the person sits directly facing the camera, framed from the waist up, centered in the frame. The camera is fixed. At second 0 the person smiles at the camera, opens both hands from the chest outward to the sides, palms up, fingers spread, wrists turning with the rhythm of speech, lasting about 2 seconds. At second 2.5 both hands are brought back in front of the chest, palms facing each other, fingers curved, as if holding a box, shaking lightly twice with the stressed beats. From second 4.5 onward both hands are placed below the lower edge of the frame, and the person only nods once lightly at the end of each short sentence.