SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsh3-prompt-writing

h3-prompt-writing

Agents & models

View repository

Updated 10-16 06:15

About this skill

H3 prompt writing (foundation)

This skill set'sgeneral foundation: only responsible for writing intent into prompts H3 can directly consume。
For styles and genres (product films, 3D, paper craft, etc.), see other skills in the same repository; after they produce drafts, this skill can compress them into standard fields.

When to use / not use

Use: writing or revising H3 prompts; image/video/audio references need tagging; cross-shot continuity; precise physics such as prop handoffs.
Don't use: copy unrelated to H3, long drama scripts, post-production engineering notes.

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, can be 2K
Frame rate/audio 24fps; stereo 32kHz
Mode Input Writing key points
T2VA Plain text Complete audiovisual timeline
I2VA 1 image first frame Anchor from the image and develop forward
FL2VA First + last 2 images Continuous path from first to last; default single-shot interpolation
L2VA 1 image last frame Reasonable prior state → converge to last frame
Ref2VA Images ≤9 videos ≤3 audio ≤3 total ≤12 Six-segment structure; tags globally synonymous

Writing process

  1. Determine mode.
  2. Read references/base-zh.txt(basic) or ref-zh.txt(full reference); English originals base-en.txt / ref-en.txt。
  3. Description too coarse → "eight-step expansion"; user already provided a complete storyboard → only check duration/timestamps/numbering/contradictions, do not compress arbitrarily.
  4. Strong continuity / multiple ordered images / precise physics → first write a Chinese storyboard plan, then compress into the final draft.
  5. Final draft: field names, shot markers, and relation markers use fixed English forms; body text defaults to English (more stable); dialogue/lyrics/on-screen text retain the original. When the user wants Chinese body text, field names remain unchanged.

Eight-step expansion

Output goal → subject and reference assets → timeline → scene → shots → visual style → sound → constraints (must keep/forbidden).
Do not fabricate brands, dialogue, or unsafe content.

Basic mode final draft

Mode instruction (first line; none for T2VA)

I2VA

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

FL2VA（N=last shot number,S.SS=duration to two decimal places)

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

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.

After a blank line, write three fields:

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...
Field What to write
integrated_multimodal_description On the timeline: visuals, actions, camera, speaker, dialogue/singing, diegetic sound
overall_soundscape Whole-film environment/physics/non-verbal vocals; 1–4 sentences; do not repeat dialogue
non_diegetic_music Music audible only to the audience (instruments, tempo, dynamics); if none, N/A

Keyframes go into the description

  • I2VA：<Picture 1>=0.00s first frame → anchor → initiation → development → result
  • FL2VA: initial state → observable intermediate changes → narrowing → final state; the end lands at the last shot's film end
  • L2VA: prior state → path → last shot aligned to image

Shot & camera movement

  • [Shot 1] no timestamp; subsequent shots [Shot N] At 00:MM.SSS, ..., cut points increase and ≤ film length
  • Hard cut:the camera cuts to etc.; dissolve only when explicitly requested by the user
  • Write camera movement into the sentence: type + amplitude/speed when needed (Zoom/Push/Pan/Truck/Tilt/Pedestal/Arc/Tracking/Static/Shake/POV/Roll;with small|large amplitude；at slow|fast speed）

Speaker and words

  • (S1) (S2) stable across shots; chorus (S1,S2)
  • Dialogue only in <d>[Language] ...</d>, original text unchanged
  • Voice-over:says in an off-screen voiceover + closed mouth on screen
  • Cross-cut <scenetrans>; end-of-film cutoff <cutoff>
  • On-screen text: wrap original text in English double quotes

Pattern skeleton

Pattern Skeleton
T2VA Style + composition → action and sound → shot cut supplements information
I2VA First frame fully locked → initiation → development → result
FL2VA Initial state → intermediate change → end state
L2VA Previous state → path → landing frame

Loose single segment: one compact segment suffices. Multiple ordered images/strong physics: use the storyboard planning below.

Full reference Ref2VA

Fixed order:

subject_definitions:
summary:
retention_analysis:
detailed_description:
overall_soundscape:
non_diegetic_music:
Label Purpose
<Subject N> Reusable visible content (not the source file itself)
<Picture N> Create separately when an image serves as a specific frame/composition anchor; if it only defines a character, write it into Subject
<Video N> Editing source, continuation starting point, whole-film structure
<Audio N> Audio to copy or reference

summary prefixes:keyframe completion / reference generation / video editing / video continuation / audio reuse / audio reference, for multiple types use +。

Frame retention markers:fully_preserved | partially_preserved | attribute_transfer | weak_reference
Audio:fully_copy | partially_copy | reference | weak_reference

detailed_description: style 1–2 sentences → shot by shot; generation task about 350–500 English words; write labels clearly on first appearance, only reuse them in later shots.
See for detailed rules and complete examples references/ref-zh.txt。

Chinese storyboard planning (then compress into final draft)

Applicable: ≥2 ordered images, cross-shot continuity, precise handoff.

  • Physically feasible within 4–15s; one active action beat per shot
  • Timestamps continuous, end point = D; default 0.5s precision
  • Two-level timeline: shot-level range + intra-shot micro-beats (establish/prepare/core/stabilize)
  • Shot N end state locked = shot N+1 start (pose, gaze, both hands, prop owner/position/connection, camera side)
  • Handoff causality: ownership → contact and firm hold → release → new ownership; do not compress into one muddled sentence

Template

【输出规格】{duration}s · {ratio} · {用途} · {N}张有序分镜图 · 不跳镜 · 锁人物服装位置道具

【整体风格】{媒介} · {视觉} · {光色纹理节奏运镜} · 场景{…} · 声音{…}

【人物与空间】人物1/2：外观与固定位置 · 硬约束{脸发型服装座位朝向…}

【道具】全片仅{名称数量} · 初态{主/手/位/连接} · 禁凭空增减跳位暗改归属

【两级时间轴】
镜头1｜0:00-{T1}｜{D1}s｜图1
运镜：…  起始：…  微轴：建立→预备→核心→保持  结束锁定：…
镜头2｜… 起始=镜头1结束锁定 …

【动作顺序】A→B→C  每镜一事  交接写清谁不动/谁拿稳/谁松手
【负面】禁身份漂移变装换位闪烁畸形多余肢物体跳变跳切字幕水印

For handoff micro-timing (earphones), split at the 0.5s level when writing: "take out → stop between the two people → release only after the other person has it firmly clipped → not inserted into ear".

After planning: write basic pattern into integrated_multimodal_description; write full reference into subject_definitions + detailed_description + retention_analysis。

Effect tips (officially synced)

  • The total described duration must align with the requested film length (4–15 seconds)
  • Reference labels consistent throughout (e.g. <Picture 1>、<Video 1>、<Audio 1>）
  • Use fewer empty words like "cinematic" and "good-looking"; write more specific audiovisual details
  • I2VA / FL2VA / L2VA: write clearly how the first frame and/or last frame connects to the timeline

Final draft self-check

  • Pattern and instruction line correct
  • Field order/label format correct
  • Time continuous, end point = duration
  • No dangling labels; states between shots reconcilable
  • Dialogue/visible text original; no brand dialogue added without authorization
  • Soundscape and score fields correctly divided
Reference Content
base-zh.txt / ref-zh.txt Chinese Guide + English Examples
base-en.txt / ref-en.txt Official English Original (consistent with upstream)