H3 Prompt Writing (Base)

This is the universal base of the skill set: it only converts intentions into prompts that H3 can directly use.
For styles and themes (product videos, 3D, paper art, etc.), refer to other skills in the same repository; after they generate drafts, you can use this skill to compress them into standard fields.

When to Use / Not to Use

Use: Writing or revising H3 prompts; tagging image/video/audio references; continuous cross-shot sequences; precise physical interactions like prop handovers.
Do Not Use: Copy unrelated to H3, long drama scripts, post-production project instructions.

Output Specifications

Item	Value
Duration	4–15s
Aspect Ratio	21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16
Resolution	Default short side 768, optional 2K
Frame Rate/Audio	24fps; Stereo 32kHz

Mode	Input	Key Writing Points
T2VA	Text-only	Complete audio-visual timeline
I2VA	1 first-frame image	Anchor from the image and develop forward
FL2VA	2 images (first + last)	Continuous path from first to last frame; default single-shot interpolation
L2VA	1 last-frame image	Reasonable previous state → converge to the last frame
Ref2VA	≤9 images, ≤3 videos, ≤3 audios, total ≤12	Six-section structure; globally synonymous tags

Writing Process

Determine the mode.
Read
```
references/base-zh.txt
```
(basic) or
```
ref-zh.txt
```
(full reference); English versions are
```
base-en.txt
```
/
```
ref-en.txt
```
.
If the description is too vague → use the "Eight-Step Expansion"; if the user has provided complete storyboards, only check duration/timestamps/numbering/contradictions, do not compress without permission.
For strong continuity / multiple ordered images / precise physics → first write a Chinese storyboard plan, then compress into the final draft.
Final Draft: Field names, shot markers, and relationship markers use fixed English expressions; the main text is in English by default (more stable); keep the original text for dialogue/lyrics/on-screen text. When the user requests Chinese main text, keep field names unchanged.

Eight-Step Expansion

Output target → Subject and reference assets → Timeline → Scene → Shot → Visual style → Sound → Constraints (required/prohibited).
Do not invent brands, dialogue, or unsafe content.

Final Draft for Basic Modes

Mode Instruction (First line; none for T2VA)

I2VA

text

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

FL2VA (

=last shot number,

S.SS

=duration with two decimal places)

text

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot N) aligns with the S.SS-second mark of the target video.

L2VA

text

How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the S.SS-second mark of the target video.

Leave a blank line and write three fields:

text

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

Field	Content to Write
`integrated_multimodal_description`	On-screen visuals, actions, shots, speakers, dialogue/singing, in-screen sounds on the timeline
`overall_soundscape`	Full-video environment/physical/non-verbal human sounds; 1–4 sentences; do not repeat dialogue
`non_diegetic_music`	Only background music audible to the audience (instruments, tempo, dynamics); write `N/A` if none

Keyframes in Description

I2VA:
```
<Picture 1>
```
= 0.00s first frame → anchor → momentum → development → result
FL2VA: Initial state → observable intermediate changes → narrowing → final state; end at the last shot's end frame
L2VA: Previous state → path → align with the last shot's frame

Shots and Camera Movements

```
[Shot 1]
```
has no timestamp; subsequent shots use
```
[Shot N] At 00:MM.SSS, ...
```
, with increasing cut points ≤ video length
Standard cuts:
```
the camera cuts to
```
etc.; cross-dissolve only if explicitly requested by the user
Write camera movements within the sentence: type + amplitude/speed if needed (Zoom/Push/Pan/Truck/Tilt/Pedestal/Arc/Tracking/Static/Shake/POV/Roll;
```
with small|large amplitude
```
;
```
at slow|fast speed
```
)

Speakers and Text

```
(S1)
```
```
(S2)
```
remain stable across shots; chorus uses
```
(S1,S2)
```
Dialogue is only placed within
```
<d>[Language] ...</d>
```
, do not rewrite the original text
Voiceover:
```
says in an off-screen voiceover
```
+ ensure the character's mouth is closed on-screen
Cross-cut:
```
<scenetrans>
```
; end truncation:
```
<cutoff>
```
On-screen text: Wrap original text in English double quotes

Mode Frameworks

Mode	Framework
T2VA	Style + composition → actions and sounds → cut shots to supplement information
I2VA	Lock first frame completely → momentum → development → result
FL2VA	Initial state → intermediate changes → final state
L2VA	Previous state → path → end at last frame

Loose single segment: A compact single segment is sufficient. For multiple ordered images/strong physics: use the storyboard plan below.

Full Reference Ref2VA

Fixed order:

text

subject_definitions:
summary:
retention_analysis:
detailed_description:
overall_soundscape:
non_diegetic_music:

Tag	Usage
`<Subject N>`	Reusable visible content (not the source file itself)
`<Picture N>`	Create separately when using images as specific frames/composition anchors; only write into Subject if defining roles
`<Video N>`	Editing source, continuation starting point, full-video structure
`<Audio N>`	Audio to copy or reference

summary

prefixes:

keyframe completion

reference generation

video editing

video continuation

audio reuse

audio reference

, use

for multiple types.

Visual retention markers:

fully_preserved

partially_preserved

attribute_transfer

weak_reference

Audio:

fully_copy

partially_copy

reference

weak_reference

detailed_description

: 1–2 sentences about style → shot-by-shot description; approximately 350–500 English words for generation tasks; clarify tags when first mentioned, reuse them in subsequent shots.
For detailed rules and complete examples, see

references/ref-zh.txt

Chinese Storyboard Plan (Then Compress into Final Draft)

Applicable for: ≥2 ordered images, continuous cross-shot sequences, precise handovers.

Physically feasible within 4–15s; one main action beat per shot
Timestamps are continuous, end point =
```
D
```
; default precision is 0.5s
Two-level Timeline: Shot-level range + micro-beats within the shot (setup/preparation/core/stabilization)
End state of Shot N = starting state of Shot N+1 (posture, gaze, hands, prop owner/location/connection, camera side)
Handover causality: Ownership → hold firmly upon contact → release → new ownership; do not compress into ambiguous sentences

Template

text

【Output Specifications】{duration}s · {ratio} · {purpose} · {N} ordered storyboard images · No shot skipping · Lock character clothing, positions, and props

【Overall Style】{medium} · {visuals} · {lighting, color, texture, rhythm, camera movement} · Scene{…} · Sound{…}

【Characters and Space】Character 1/2: Appearance and fixed position · Hard constraints{face, hairstyle, clothing, seat, orientation…}

【Props】Only {name, quantity} in the whole video · Initial state{owner/hand/location/connection} · 禁凭空增减跳位暗改归属

【Two-level Timeline】
Shot 1｜0:00-{T1}｜{D1}s｜Image 1
Camera Movement:…  Start:…  Micro-axis: Setup→Preparation→Core→Hold  End Lock:…
Shot 2｜… Start = Shot 1 End Lock …

【Action Sequence】A→B→C  One action per shot  Clearly write who stays still/who holds firmly/who releases during handover
【Negative Constraints】Forbid identity drift, costume changes, position shifts, flickering, deformities, extra limbs, object jumps, jump cuts, subtitles, watermarks

For micro-timing of handovers (e.g., headphones), split into 0.5s steps during writing: "Take out → stop between two people → release only after the other party holds firmly → not inserted into ear".

After planning: For basic modes, write into

integrated_multimodal_description

; for full references, write into

subject_definitions

detailed_description

retention_analysis

Final Draft Self-Check

Correct mode and instruction line
Correct field order/tag format
Continuous timeline, end point = duration
No dangling tags; inter-shot states are reconcilable
Original dialogue/visible text preserved; no unauthorized brands or dialogue added
Correct division of labor between soundscape and music fields

Reference	Content
`base-zh.txt` / `ref-zh.txt`	Chinese guide + English examples
`base-en.txt` / `ref-en.txt`	Official English original

h3-prompt-writing

NPX Install

Tags

SKILL.md Content (Chinese)

H3 Prompt Writing (Base)

When to Use / Not to Use

Output Specifications

Writing Process

Eight-Step Expansion

Final Draft for Basic Modes

Mode Instruction (First line; none for T2VA)

Keyframes in Description

Shots and Camera Movements

Speakers and Text

Mode Frameworks

Full Reference Ref2VA

Chinese Storyboard Plan (Then Compress into Final Draft)

Template

Final Draft Self-Check