Menu
Generate video content using Gemini Veo. Use when user asks to create video, TikTok ad, Instagram Reel, YouTube video, product video, explainer.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
MANDATORY pre-flight check that runs BEFORE any content plan or generation. Classifies the topic, verifies project/brand/product context, resolves the required reusable assets (characters, products, logos, voices) from the asset registry, and provides/generates whatever is missing — saving paths for reuse. Use whenever the user asks to plan a campaign, make a marketing plan, or create/generate any image, video, voiceover, or music.
Generate music using Gemini Lyria. Use for background music, jingles, intros, soundtracks, and complete songs WITH sung vocals — Lyria 3 sings your exact lyrics (8 languages, rap/sung delivery).
Generate images using Gemini 3 Image (Nano Banana). Use for thumbnails, carousels, infographics, social graphics, product photos.
Generate voiceover audio using Gemini TTS. Use for narration, voiceovers, podcasts, multi-speaker dialogue.
Write the words — scroll-stopping hooks, framework-driven scripts, and platform-ready captions with hashtags. Use when the user needs hook ideas, a video/ad script, ad copy, post captions, or says "what should I post", "write a script", "give me hooks", "write the caption".
Analyze a reference video (YouTube URL or local file) into a scene-by-scene breakdown — timecodes, shot types, camera moves, on-screen text, spoken lines — plus a recreation blueprint with per-scene video prompts and VO script. Use when the user says "analyze this video", "I want to create something like that", "recreate this video", "break down this YouTube video", or pastes a video link they want to emulate.
| name | generate-video |
| description | Generate video content using Gemini Veo. Use when user asks to create video, TikTok ad, Instagram Reel, YouTube video, product video, explainer. |
| allowed-tools | Bash Read Write Edit Glob Grep |
You MUST complete ALL steps below and SHOW output to user before generating.
□ Step 0: RUN content-preflight (classify topic/visual mode, resolve required assets)
□ Step 1: ASK user "Does this video include a person/character?"
□ Step 2: Look up the character/product in the asset registry (resolveAsset) — SHOW result
□ Step 3: If a locked character asset exists → TELL user you will reuse its ref files
□ Step 4: READ project.md and brand.md templates
□ Step 5: ASK user to confirm workflow selection
□ Step 5b: Any character SPEAKS? → WRITE THE EXACT LINES in the prompt, and add NO voiceover
□ Step 6: THEN generate using workflow function
DO NOT skip to code. DO NOT assume. SHOW each step.
Applies to all Omni modes (text_to_video, image_to_video,
reference_to_video) and to Veo/Seedance speaking clips.
The model always generates the speech. If you don't write the lines, it invents
them — and its invention can reverse your story. Never write "speaks a short
reassurance" or "mouth moving in natural speech". Write says: <exact words>.
Then do NOT layer a TTS voiceover on that scene. The clip already carries its
dialogue. Voiceover (generateVoiceover) is for narration only — a narrator
over visuals with no on-screen speaker.
In assembleStoryFilm: speaking scenes keep ambientVolume: 1.0 and pass no
voPath. Only narration scenes get a voPath.
Full rules + the production incident that produced them:
workflows/VIDEO-PROMPT-GUIDE.md §4e / §4e-vo.
DO NOT write raw API calls. DO NOT author a .cjs/.ts file that imports these
functions. RUN the workflow as a CLI command:
node workflows/cli.cjs <command> '<json-args>' # or @args.json for long prompts
Video commands (run node workflows/cli.cjs list for all):
| Command | Use for |
|---|---|
generateSilentVideo | B-roll / product video, no voice |
generateSpeakingVideo | On-camera character speaking |
generateVideoFromImage | Animate an approved keyframe |
generateSpeakingVideoFromImage | Character image → speaking clip |
generateVideoWithVoiceover | Video + generated voiceover |
generateVideoFromImageWithVoiceover | Keyframe → video + voiceover |
generateOmniVideoClip | Omni Flash: stylized explainers, text-in-scene, multi-ref consistency, edit existing clips |
Example:
node workflows/cli.cjs generateSilentVideo \
'{"prompt":"...","outputPath":"projects/{name}/output-contents/clip01.mp4","duration":8,"aspectRatio":"9:16","quality":"fast"}'
For each command's argument shape: Read workflows/WORKFLOWS.md. The import
signatures there document the args — translate them into a cli.cjs call.
Storyboard before you pay for clips. For any multi-scene/character video, generate a keyframe per scene with
generateStoryboard()(~$0.067 each), get the user's approval, then generate clips from the approved keyframes. See thepreview-pickskill. Need the script/hook first? Use thewrite-copyskill (generateHooks→generateScript).
ALWAYS ask the user:
Does this video include a person/character?
If yes, ask:
Reference material? (ask once, before writing any video prompt)
analyze-video skill FIRST and build from its
recreation blueprint instead of improvising prompts.style_references in the registry (resolveAsset) — registered look
refs feed keyframe prompts and referenceImagePaths ("style only, not subject").Don't guess from filenames. Read the project's asset registry — it records every
reusable character/product/voice with validated paths. (Pre-flight has usually populated
it already; see skills/content-preflight/SKILL.md.)
node workflows/cli.cjs loadAssetConfig '["{name}"]'
node workflows/cli.cjs resolveAsset '["{name}","char-main"]' # the character id from the content plan
node workflows/cli.cjs resolveAsset '["{name}","prod-main"]'
# resolveAsset → { ok, existing[], missing[] }
Then TELL the user what you found:
Asset registry (projects/{name}/config/assets.yaml):
✓ character char-main → 3 refs, locked (will reuse for consistency)
✓ product prod-main → 2 refs
⚠ no voice linked to char-main
Reusing char-main's reference images — not regenerating the character.
char.ok === true → pass char.existing[...] as the reference image(s). Do not
re-describe or regenerate a locked character.char.ok === false (missing or no asset) → resolve it via pre-flight's
provide/generate flow, then registerAsset(...) so it's saved for next time.cat projects/{name}/templates/project.md
cat projects/{name}/templates/brand.md
cat workflows/WORKFLOWS.md
Problem: Each video generation creates a NEW random character.
Solution: Use the SAME character reference image for ALL video clips.
Resolve the character from the registry:
node workflows/cli.cjs loadAssetConfig '["{name}"]'
node workflows/cli.cjs resolveAsset '["{name}","char-main"]' # → { ok, existing[], missing[] }
If char.ok → reuse the SAME ref file for every clip:
# reuse char.existing[0] from resolveAsset (the validated, on-disk path) as referenceImagePath
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"<char.existing[0]>","dialogue":"Dialogue for this clip"}' # ...more args
If not (no asset yet) → generate ONCE, then register it:
# Step 1: Generate the character keyframe one time
node workflows/cli.cjs generateSingleImage '{"prompt":"Woman in her 30s, dark hair, professional attire, neutral expression","outputPath":"projects/{name}/assets/characters/char-main-front.png","aspectRatio":"9:16"}'
# Step 2: Save it to the registry so EVERY future clip/piece reuses it
node workflows/cli.cjs registerAsset '["{name}","characters",{"id":"char-main","label":"Main character","description":"Woman in her 30s, dark hair, professional attire","source":"generated","status":"ready","locked":true,"files":{"front":"assets/characters/char-main-front.png"}},{"date":"<today ISO>"}]'
# Step 3: Use it as the reference for all videos
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"projects/{name}/assets/characters/char-main-front.png"}' # ...more args
For multi-clip videos / series:
registerAsset(... locked:true)files entries under the same id when neededDoes this video include a person/character?
generateSpeakingVideo() or generateSpeakingVideoFromImage()generateVideoFromImage() with character referencegenerateSilentVideo() or generateVideoFromImage() with productPlatform? (TikTok, Instagram, YouTube)
Duration? (15s, 30s, 60s)
Audio style? — ALWAYS ask, don't assume:
generateSilentVideo() or generateVideoFromImage()generateSpeakingVideo() / generateSpeakingVideoFromImage()generateVideoWithVoiceover() /
generateVideoFromImageWithVoiceover()If voiceover is chosen, do NOT auto-pick the voice. Run the voiceover
suggest-and-confirm flow (skills/generate-voiceover/SKILL.md Steps 1c–3): propose
voice / style / pace / accent / script / pronunciation / music-under-VO with reasoning,
show the summary, and get the user's confirmation before generating. Reuse the
project's locked voice (or the on-screen character's linked_voice) if one exists, but
still confirm it.
Reference images? (check assets folder)
If generating with Omni Flash — ART STYLE (REQUIRED, do not skip): Present the preset list and ask the user to pick ONE:
Omni Flash art style — which look?
1. photorealistic (default — no preset)
2. pixel-art 7. low-poly
3. claymation 8. 3d-mix
4. mixed-media 9. isometric-flat-vector
5. 3d-papercraft 10. fluffy-toy
6. whiteboard-doodle 11. 2d-illustration
…or describe your own custom style
Pass the chosen id as "artStyle" (omit for photorealistic). Record the
choice in projects/{name}/templates/brand.md so the whole series stays
in one style. Never silently pick a style for the user.
Camera movement — REQUIRED for EVERY clip/scene. Check the guide, don't improvise:
A clip with no camera direction renders as a locked-off static shot — dead
energy. For EVERY clip in the piece, READ VIDEO-PROMPT-GUIDE.md §2b (46
presets) and pick 1-2 ids that fit that clip's intent, tell the user which
you chose and why, and confirm. One move per clip; vary moves across
consecutive clips (two identical moves back-to-back reads as a loop).
static is itself a deliberate preset — allowed only when the user
explicitly wants a locked-off look, never as a default by omission.
| Intent | Propose |
|---|---|
| Product hero / reveal | slow-zoom-in, orbit-clockwise, dolly-in |
| Calm premium B-roll | static, slider-right, slow-zoom-out |
| Hype / hook energy | crash-zoom-in, whip-pan-right, chase |
| UGC / authentic | handheld, first-person |
| Walking testimonial | reverse-tracking, side-tracking |
| Location / scale reveal | drone-pull-back, crane-up, earth-zoom-out |
| Scene transition | whip-pan-*, pass-through, infinite-zoom |
Pass it as "cameraMove":"<id>" on generateSilentVideo,
generateVideoFromImage, or generateOmniVideoClip — the full
Movement/Speed/Framing/End block is prepended automatically. For
scene-array workflows (voiceover scenes, speaking video, Seedance), paste
the block text from the guide at the head of each scene prompt instead.
Series/film/campaign? → Style Block (one locked look on every prompt):
For a cinematic film, animated series, or product-shot campaign, write
projects/{name}/templates/style-block.md ONCE with the user (template +
variants: templates/style-block.template.md), register it locked
(style_references), then prepend the block verbatim to every clip,
keyframe, and image prompt of that series — note "style-block: v1" in
prompts.txt. The block owns the LOOK (medium, lighting, color, lens);
cameraMove owns the MOVEMENT (question 7) — never put camera movement
inside the block. Skip for one-off clips; brand.md aesthetic is enough there.
| User Wants | Has Person? | Has Reference? | Audio | Use Workflow |
|---|---|---|---|---|
| B-roll, ambient | No | No | Silent | generateSilentVideo() |
| Product video | No | YES | Silent | generateVideoFromImage() |
| UGC, testimonial | YES | No | Character speaks | generateSpeakingVideo() |
| Character video | YES | YES | Character speaks | generateSpeakingVideoFromImage() |
| Explainer | No | No | Voiceover (TTS) | generateVideoWithVoiceover() |
| Product explainer | No | YES | Voiceover (TTS) | generateVideoFromImageWithVoiceover() |
RULE: If video has person AND character image exists → ALWAYS use character image as reference.
node workflows/cli.cjs generateSilentVideo '{"prompt":"Cinematic shot of smartwatch on wrist, golden hour","outputPath":"projects/my-project/output-contents/video.mp4","duration":6,"aspectRatio":"9:16","quality":"fast"}'
node workflows/cli.cjs generateVideoFromImage '{"referenceImagePath":"projects/my-project/assets/product-watch.png","prompt":"Product rotates slowly, premium lighting","outputPath":"projects/my-project/output-contents/product.mp4","duration":6,"aspectRatio":"9:16","cameraMove":"orbit-clockwise"}'
cameraMove (optional, also on generateSilentVideo / generateOmniVideoClip)
prepends a tested four-part camera block — 46 preset ids in
workflows/VIDEO-PROMPT-GUIDE.md §2b (see Required Question 7).
# Use character image from assets for consistency
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"projects/my-project/assets/characters/char-main-front.png","environment":"Modern office, soft lighting","dialogue":"This watch literally saved my life.","voiceDescription":"warm female voice, early 30s","outputPath":"projects/my-project/output-contents/testimonial.mp4","duration":8,"aspectRatio":"9:16"}'
# Only use when NO character image exists
node workflows/cli.cjs generateSpeakingVideo '{"characterDescription":"Woman in her 30s, dark hair, professional","environment":"Modern office, soft lighting","dialogue":"This watch literally saved my life.","voiceDescription":"warm female voice, early 30s","outputPath":"projects/my-project/output-contents/testimonial.mp4","duration":8,"aspectRatio":"9:16"}'
# Step 1: Resolve the locked character once from the registry — use .existing[0] as the reference below
node workflows/cli.cjs loadAssetConfig '["my-project"]'
node workflows/cli.cjs resolveAsset '["my-project","char-main"]'
# Step 2: Generate all clips with the SAME reference (char-main .existing[0])
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"<char-main existing[0]>","dialogue":"First dialogue...","outputPath":"projects/my-project/output-contents/clip-01.mp4","duration":5}'
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"<char-main existing[0]>","dialogue":"Second dialogue...","outputPath":"projects/my-project/output-contents/clip-02.mp4","duration":5}' # SAME reference
node workflows/cli.cjs generateSpeakingVideoFromImage '{"referenceImagePath":"<char-main existing[0]>","dialogue":"Third dialogue...","outputPath":"projects/my-project/output-contents/clip-03.mp4","duration":5}' # SAME reference
Two different mechanisms on generateVideoFromImage:
referenceImagePath — first frame: the clip starts from this exact image
(animate an approved keyframe).referenceImagePaths — asset references, max 3 (Veo 3.1): character
sheet + environment sheet + prop in ONE request; they guide what things look
like without appearing verbatim. State each ref's role in the prompt
("the girl from the character sheet", "the garden from the environment
reference"). Need 4-5 refs → generateOmniVideoClip (<IMG_REF_n> tags).When picking refs for a scene, prioritize: character sheet (if the scene has a person) → prop/product sheet → environment sheet. Drop the environment ref first if over the limit — environments survive prompt description better than faces do.
❌ NEVER include:
✅ Only describe visuals:
Text is added via FFmpeg post-production.
| Quality | Cost/sec |
|---|---|
| lite | $0.03 |
| fast | $0.08 |
| standard | $0.20 |
Example: 20s video at 'fast' = $1.60
Check the budget cap BEFORE generating (the CLI hard-stops at BUDGET_EXCEEDED):
node workflows/cli.cjs checkBudget '["{name}", 1.60]' # project name + estimated cost
<content-id>.pipeline.json IN the content folder (nodes = CLI commands,
{{node.data.field}} refs wire outputs to inputs), then
node workflows/cli.cjs runPipeline @<file>. See workflows/pipelines/README.md.createGenerationManifest /
addManifestEntry CLI commands (AGENT-GUIDE Step 5).Professional: Kore, Charon, Orus, Fenrir Friendly: Zephyr, Puck, Aoede, Leda Energetic: Calliope, Proteus Calm: Autonoe, Despina, Aura
After the clips/audio exist, assemble the deliverable with the assembly workflows (local ffmpeg, no extra API cost). Ask the user whether they want burned-in captions — short-form (TikTok/Reels/Shorts) usually should.
# 1. (optional) Captions from the voiceover/dialogue script
node workflows/cli.cjs generateCaptions '{"script":"<voiceover/dialogue script>","totalDuration":30,"outputPath":"projects/{name}/output-contents/captions.srt"}'
# 2. One call: concat clips → lay voiceover + music (music auto-ducks) → burn captions
node workflows/cli.cjs assembleFinal '{"clipPaths":["…/clip-01.mp4","…/clip-02.mp4"],"voiceoverPath":"…/voiceover.wav","musicPath":"…/music.wav","musicVolume":0.3,"captionsSrtPath":"…/captions.srt","outputPath":"projects/{name}/output-contents/final.mp4"}'
# voiceoverPath / musicPath / captionsSrtPath are optional; omit captionsSrtPath to skip burned-in captions; musicPath reuses the registry music bed
# Optional between-clip transitions: add "transition":"dissolve" + "transitionDuration":0.5
# 58 viable presets (pick by intent): slide=slideleft/right/up/down · swipe=wipe*/smooth* ·
# cover*/reveal* · zoom=zoomin/squeezeh/squeezev · fades=fade/fadeblack/fadewhite/dissolve ·
# stylized=pixelize/distance/hblur/*wind · shapes=circleopen/radial/diagtl… Full table: VIDEO-PROMPT-GUIDE.md § Assembly transitions.
# NOT YET (guard rejects): glitch, roll, zoomout. Hard cuts when transition omitted.
# Each overlap shortens the total by its duration; every clip must outlast it.
For audio-only mixing (single clip + VO/music, no concat/captions) use mixVideoAudio().
| Route | Cost | Command | Use for |
|---|---|---|---|
| A — Normal captions | $0 | renderCaptionedVideo / TextMotion scrim:true | transcript pills, white-on-dark at y=0.75, accent keywords — every VO video |
| B — Creative text, AI-designed | $0 | renderTextMotion (+ rembg behind-subject) | hero/kinetic words the AGENT designs: frame-grab → place in the SUBJECT's negative space (close 9:16 talking head → band above the head, y≈0.09 — never on a face) → words from the video's own transcript → per-word style/motion → frame-grab verify BEFORE assembly |
| C — Creative text, Omni-baked | PAID ≈$0.10/sec | generateOmniVideoClip task edit | text drawn/painted INTO the scene. Timestamped beats (00:02 — {…}) work; keep prompts minimal; behind-subject occlusion works ONLY at head height; output silent → mixVideoAudio; 10s cap → split+concat+setpts. Full rules: TEXT-OVERLAY-DESIGN-GUIDE.md §8. Confirm spend first. |
⚠️ For CREATIVE text, ASK the user first — "free (Remotion, agent-designed) or
paid with Gemini Omni (~$0.10/sec, in-scene drawn look)?" — recommend free, never
route to Omni silently. If they pick Omni, CALCULATE the job's total (count every
≤10s call the scenes need × ~$0.10/sec, e.g. 18s scene = 2 calls ≈ $1.90), state
it with the budget position, and checkBudget before the first call.
Full route selector + house styles + the ask script: workflows/TEXT-OVERLAY-DESIGN-GUIDE.md §0-pre.
assembleFinal's captionsSrtPath burns a plain SRT. For anything more expressive, run
renderCaptionedVideo on the finished video — and decide the method from the video
type (a talking-head, a hook reel, and a multi-scene travel piece want different
caption treatments):
workflows/TEXT-OVERLAY-DESIGN-GUIDE.md § 0 (caption
method selector). E.g. talking-head → bottom pill transcript + **keyword** accent;
hook/sizzle → style:"hero" word punch-ins; multi-scene → upper location stamp +
transcript; signature "wow" moment → text-behind-subject.style pill/hero · pos upper/mid/lower · size · color ·
**word** accent · \n stacked lines): WORKFLOWS.md § renderCaptionedVideo. Timing
from transcribeAudio, offset by each clip's start on the assembled timeline. Only
one cue is active at a time → stack layers (transcript + stamp) with separate passes.rembg per-frame matte):
recipe in TEXT-OVERLAY-DESIGN-GUIDE.md § 6 — bake into each clip BEFORE assembly, then
add the transcript pass. Needs the rembg lib — check/install first:
python3 -c "import rembg" || python3 -m pip install "rembg[cpu]" pillow (first run
auto-downloads a ~176 MB model to ~/.u2net/).Save to: projects/{name}/output-contents/{date}/
final.mp4 (assembled: clips + voiceover + music + optional captions)video.mp4 or final-with-voiceover.mp4clip-01.mp4, clip-02.mp4, etc.captions.srt (if generated)| Platform | Aspect | Duration |
|---|---|---|
| TikTok | 9:16 | 15-60s |
| Instagram Reels | 9:16 | 15-90s |
| YouTube Shorts | 9:16 | 15-60s |
| YouTube | 16:9 | 60s+ |
For word-driven reels (quotes, stats, hooks over brand backgrounds), skip Veo:
generate text-free backgrounds + voiceover, then one call renders staggered
animated typography (per-line size/color emphasis, gold #C8A24A for THE phrase):
node workflows/cli.cjs renderKineticReel '{"scenes":[…],"audioPath":"vo.wav","outputPath":"reel.mp4"}'
Args in WORKFLOWS.md § Remotion. Use Veo only when you need real motion/footage.
For per-element creative control over ANY footage (Route B above) use
renderTextMotion: per-word font/size/color/gradient/stroke/glow/badge,
20+ entrances, loops (float/pulse/wave/neon/wiggle…), letter/word stagger,
mediaFill (video-inside-letters), behind:true (rembg subject matte),
scrim:true caption pills, background.playbackRate to stretch a short
background. Design rules + verification steps: TEXT-OVERLAY-DESIGN-GUIDE.md §0-pre.
Before planning ANY hero title / kinetic reel / "text behind subject" /
meme-style overlay copy, read workflows/TEXT-OVERLAY-DESIGN-GUIDE.md —
placement (behind/front/side), size-hierarchy patterns (hero+support,
hook→context→CTA, question→answer reveal), color/font pairing, and —
important — its capability matrix of what's actually supported by
renderKineticReel/renderSlideStill today vs. what's roadmap-only. Don't
promise a per-word color/size effect or a rotated side badge as if it just
works; the guide names the current workaround or gap explicitly. Myanmar/
Burmese text in ANY of these must go through Remotion, never raw ffmpeg drawtext — same guide, § 7.
workflows/recipes/cinematic-story-film.md: storyline arc, character sheet →
keyframes → Veo 3.1 clips → adaptive per-scene VO → one-call
assembleStoryFilm. ~$3.65 at fast tier for 4×8s.workflows/recipes/story-short-film.md — the production-sheet pipeline:
story.md: storyline, characters+personality, locations,
items, shot list) → user approvalgenerateVideoFromImage referenceImagePaths (≤3 sheets/scene;
4-5 refs or stylized → Omni Flash)assembleFinal with transition (dissolve/fade) + VO/music/captionsgenerateVideoFromImage etc.): cinematic fidelity, 1080p+,
8s beats, animating locked NBP keyframes. $0.10/s fast tier.generateOmniVideoClip): instruction-precision + styling +
editing. ~$1.03/10s, max 10s/clip, 720p only, 16:9 or 9:16, native audio
(VO/SFX prompted in text — no audio input). Timestamped prompts
([00:00-00:02]…, He says,"…", SFX:, (no subtitles)) work on BOTH
models — see workflows/VIDEO-PROMPT-GUIDE.md § Gemini Omni Flash for the
full guide and tested verdicts.| User wants | Pass | Task |
|---|---|---|
| Explainer / sizzle reel from text | prompt only | text_to_video |
| Animate a keyframe / scene image (explainer, cinematic) | referenceImagePath | image_to_video |
| Product/character consistency in a new scene | referenceImagePaths (2-5) + cite <IMG_REF_0>… in prompt | reference_to_video |
| Add SFX / on-video text / restyle / camera change on an existing clip | inputVideoPath + one-change instruction | edit |
| Motion control — swap the performers in a real video with your characters | inputVideoPath + referenceImagePaths + minimal replace-prompt | edit |
Replaces the people in a real video (dance, choreography, any performance)
with your locked characters; motion, timing, camera, and background carry
over 1:1. Full recipe with receipts: workflows/VIDEO-PROMPT-GUIDE.md §7.
The rules that change outcomes:
<IMG_REF_0>,
<IMG_REF_1>, and <IMG_REF_2> (left, center, right respectively),
keeping the same choreography, timing, camera, background, and lighting.
(no subtitles)". Descriptive breed/outfit prompts get Input-blocked
($0 each) — don't reword adjectives, strip them.IMG_REF_n). Outfit variants of one locked character work well via
generateImageVariation (~$0.13/still) — clothing must be text/logo-free
or it renders garbled.mixVideoAudio (musicVolume 1.0, $0,
beat-perfect since motion timing is copied 1:1).reviewVideoOutput (frameCount: 4) — watch for mid-clip
plush/mascot style drift; re-roll only if publishing.Before any stylized Omni generation: ask the art-style question (Required Question 6 above). Prefer Omni Flash over Veo for: explainers in a preset art style, live writing/text-in-scene, editing existing clips, and multi-asset consistency. Prefer Veo for: premium cinematic quality and >10s beats. The NO-TEXT-IN-PROMPTS rule is Veo-only — Omni renders and syncs text well.
Full playbook: workflows/VIDEO-PROMPT-GUIDE.md § Production-Tested Playbook.
The short version that changes decisions:
char-x-hero.png) — video models blend multi-view object
sheets into segmented toys.generateImageVariation keyframe (sheets as refs, target aspect,
~$0.07) → show/QA the STILL → animate that exact file with
generateOmniVideoClip (image_to_video) or generateVideoFromImage.
Identity lives in the keyframe; the video model only adds motion + speech.generateVideoFromKeyframes (firstFramePath + lastFramePath, Veo 3.1).| Dialogue language | Engine |
|---|---|
| English | Veo or Omni |
| Myanmar / Thai / non-Latin scripts | Omni Flash (image_to_video) — better pronunciation, no filter drama |
| Any language + lip-sync UGC realism | Seedance (needs OPENROUTER_API_KEY) |
| PROVIDED audio file (voice clone/recording) | infiniteTalkLipsync (RunPod) — the ONLY working path; audio drives the mouth, no duration cap. Omni is policy-blocked for this (deepfake guard) — do not retry it. Recipe: VIDEO-PROMPT-GUIDE §8 |
Veo + non-English script = silent filter blocks ($0, "No video was generated") unless the ENTIRE prompt is in the target language — which then weakens reference-following. Never write singing/music/voice-direction meta words in Veo prompts (instant block); prevent sing-song delivery with "asks casually in everyday spoken , plain conversational tone".
generateOmniVideoClip {inputVideoPath, referenceImagePaths:[<character sheets>], prompt} —
open with "Keep everything the same — same motion, camera, timing, audio and
spoken dialogue. Only ." Any edit touching a character MUST attach
that character's reference images — text-only edits redraw identity. Omit
duration/aspectRatio (inherited from input video).