跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

video-prompting

Base craft for AI video prompts inside Palmier Pro across Seedance, Kling, MiniMax H3/Hailuo, Grok, and Veo. Assigns @Image/@Video/@Audio authority, locks identity and screen geography, then stages timed shots. Use for generate_video multimodal prompts, precise video edits, or cinematic AI video beyond simple UGC talking-head beats. For UGC selfie-then-animate prefer ugc-photo-prompts then ugc-video-prompts. Specialized siblings: character-pipeline, cinematic-motion, motion-graphics, podcast-ad.

palmier-io · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add palmier-io/palmier-skills --skill video-prompting

技能说明


name: video-prompting description: Base craft for AI video prompts inside Palmier Pro across Seedance, Kling, MiniMax H3/Hailuo, Grok, and Veo. Assigns @Image/@Video/@Audio authority, locks identity and screen geography, then stages timed shots. Use for generate_video multimodal prompts, precise video edits, or cinematic AI video beyond simple UGC talking-head beats. For UGC selfie-then-animate prefer ugc-photo-prompts then ugc-video-prompts. Specialized siblings: character-pipeline, cinematic-motion, motion-graphics, podcast-ad.

Video Prompting

Write a production brief, not a vibe paragraph. Models fail when every asset competes for identity, motion, camera, and style.

Core rule

Assign authority before action. Every reference gets one job and an explicit boundary on what it must not control. Then: lock continuity → place screen geography → time the shot → light/sound → short negatives.

When to use which skill

NeedSkill
Multimodal / cinematic / edit promptsthis skill
Face consistency across stills → videocharacter-pipeline
Documentary handheld / density mapscinematic-motion
Overheard two-person podcast adpodcast-ad
UGC talking-head / product selfieugc-photo-prompts → ugc-video-prompts
Assemble / caption finished clipsugc-editing
Kinetic titles / flat glass UI / H3 product journey promosmotion-graphics

Session gates

  1. get_timeline — if canGenerate is false, ask user to sign in / subscribe.
  2. list_models for image / video / audio — ids and caps drift.
  3. Propose prompt + model + duration + aspect (+ refs). Wait for confirmation — paid, not undoable.
  4. Fire generate_*; do not busy-poll. Ready when get_media({ ids }) has no generationStatus. On failed, report and ask before retry.
  5. Organize with folder (e.g. Hero/Sheets, Hero/Takes).

Reference wiring

Tool argumentPrompt tagsTypical job
startFrameMediaRef / endFrameMediaRefimplied first/last frameImage-to-video anchors
referenceImageMediaRefs@Image1… (some Kling: @Element1)Identity, wardrobe, product, scene
referenceVideoMediaRefs@Video1…Motion timing, camera rhythm
referenceAudioMediaRefs@Audio1…Voice, lip-sync audio, music timing
generate_image → referenceMediaRefsmodel-specificStill identity lock

Declare authority when multiple refs are passed:

@Image1 defines identity and wardrobe only. Ignore its background.
@Video1 defines motion timing and camera only. Do not inherit its people or location.
@Audio1 defines <Character>'s voice timbre and delivery only.

Never: “use all references to make the same scene.”

Prompt architecture

Stable truth first; time second; polish last.

[REFERENCE USE]     — what each @ImageN / @VideoN / @AudioN controls + ignores
[IDENTITY LOCKS]    — count, face, hair, wardrobe, props, voice ownership
[SCENE]             — place, time of day, dramatic goal, emotional turn
[DIALOGUE]          — exact quoted lines, speakers, delivery + voice pins
[SCREEN GEOGRAPHY]  — who/what is left/right/front before anyone moves
[SHOT LIST]         — time ranges: framing, action, dialogue, end state
[ACTING]            — posture, gaze, gesture, emotional progression
[LIGHT / IMAGE]     — light, palette, texture, lens, DoF
[CAMERA]            — sizes, move rules, axis, transitions
[SOUND]             — ambience, dialogue clarity, music rule
[NEGATIVES]         — only the failure modes that matter

Length vs start frame: with startFrameMediaRef alone and simple motion, keep the prompt short (camera + action + speech/SFX) — do not re-describe the still. With multi-ref / timed performance / edits, use the full architecture.

Continuity + shot timing

Lock subject count, face/hair/wardrobe, prop ownership, screen direction, voice ownership. Place people before they move.

Treat duration as an editorial budget (list_models → durations). One primary beat per range: establish → escalate → resolve. Prefer one beat per generate_video call; assemble on the timeline. If dialogue rushes, shorten lines — do not cram two turns into one short clip.

Model routing

Confirm ids/caps via list_models before firing. Default stack for these skills:

StagePreferHow
Iterateseedance-2-mini, MiniMax H3 (hailuo-03), or Flux 3 draftMini / H3 as normal generates. Flux 3: draft: true (720p preview; supportsDraft must be true).
Finalseedance-2, MiniMax H3 (hailuo-03), or FLUX EnhanceSeedance / H3: full-quality generate (ask before Seedance 1080p/4K). Flux 3 finals are not upscale_media — only enhance a completed draft.

Flux 3 draft → enhance (not upscale)

  1. Iterate: generate_video({ model: "flux-3", draft: true, … })
  2. When get_media shows the draft ready and canEnhanceDraft: true, final: generate_video({ enhanceDraftMediaRef: "<draftId>" }) alone — no prompt/model/inputs. Keeps motion; renders FLUX.3 at 1080p.
  3. Do not use upscale_media for this path. Enhance only works on Flux 3 drafts (canEnhanceDraft); other models have no enhance step.
Other goalsPreferNotes
Synced spoken dialogueKling V3/O3, or GrokQuoted line; pin voice every beat
Stills / sheetsgrok-imagine, gpt-image-2, or seedream-v5-proBest of multi-room podcast bakeoff; see character-pipeline / podcast-ad
Avoid as defaultVeoContent-checker failures common

Precise editing

  1. Master — source / @Video1 owns timeline, subjects, camera, audio.
  2. Changes — exact subject/object/region/line/time only.
  3. Preserve — identity, occlusion, eyelines, camera, dialogue, ambience, duration.

Text and titles

Most video models cannot render reliable readable t

阅读完整来源 ↗