vox-explainer
GitHub将品牌角度转化为Vox风格的纸拼贴解说广告。支持叙事节拍图、分层拼贴关键帧生成、动态效果及配音配乐，强调产品真实照片与手绘拼贴风格的结合，适用于各类拼贴动画需求。
触发场景
安装
npx skills add NeverSight/learn-skills.dev --skill vox-explainer -g -y
SKILL.md
Frontmatter
{
"name": "vox-explainer",
"description": "Turn a brand angle into a finished Vox-style paper-collage explainer ad on the Advibly MCP. Approve a narrative beat map and visual theme, generate layered collage-poster keyframes, animate living paper motion with Gemini Omni Flash, then compose narration, music, and SFX. Supports baked headlines, photoreal product cutouts, assemble-from-empty reveals, confetti, impact shake, and editorial collage motion. Trigger for \"Vox style video\", \"paper collage animation\", \"motion collage\", \"collage explainer\", \"collage ad\", \"paper collage explainer\", \"make a collage ad\", \"explainer video for my product\", or a reference with the editorial torn-paper look. Use even without the word \"skill\" when that style is clearly requested."
}
Advibly Vox Explainer
Turn one brand angle into a finished Vox-style paper-collage explainer ad: a bold, punchy, narrated spot where each beat is a torn-paper collage poster that comes alive, cut every 4 to 6 seconds, with the real product photo living photoreal inside the paper world. Everything generates and assembles on the Advibly MCP.
The look is the modern editorial paper collage popularized by Vox explainers and creators like Stav Zilber and rom1trs: hand-cut paper cut-outs, torn edges, tape, halftone dots, newspaper clippings, bold flat color per beat, big cut-out headlines baked into the image.
Speak in the user's language. No em dashes anywhere in output; use periods or line breaks. Keep labels and on-screen copy free of emoji unless asked.
The core idea (read this first)
The collage look and the collage motion are two different steps:
- The look is born in the IMAGE step. Each shot is a finished collage poster made by the image model. All the collage DNA (torn paper, cut-outs, halftone, bold flat color, headline text) lives in that image. If the poster is not a rich layered collage, nothing downstream will save it. Re-roll cheap here rather than paying to animate a weak image.
- The motion is added after. The video model animates the whole poster into a "living collage": layers parallax, elements bob and scatter, edges flutter. The more clearly layered the poster (distinct cut-outs, visible edges, own drop shadows), the more layered motion the video model can produce. A flat blended image can only be panned as one plane.
- The Advibly twist: the product stays photoreal. Everything in the frame is printed paper EXCEPT the real product photo, which is composited in unchanged. Real product in a craft world is what sells it as an ad. This is the one element you never stylize.
Two layers drive everything, each with its own reference file. Read both before writing the beat map or any prompt:
- STORY layer:
references/beat-layer.md. Narrative arcs, hook patterns, beat counts, shot sizes, camera moves, element motion, anti-monotony rules. - LOOK layer:
references/prompt-guide.md. The image and video prompt structures, the vocabulary banks that fill them, and the theme presets.
Hard defaults (do not drift)
- Image model:
advibly_generate_imagewithmodel: "nano-banana-2",resolution: "2K". The whole Vox look lives in the paper grain, and nano-banana-2 holds that grain the best while still rendering baked headline text cleanly and taking the product photo as a reference. It is the default.gpt-image-2(quality: "high") is the fallback for one poster whose headline keeps degrading (strongest at dense baked text) or a product-reference edit that fights the collage;seedream-5-prois the option for a type-heavy poster. The Phase 3 bake-off decides the winner by eye, and that one model renders the whole ad. If a set comes back digital-smooth and plasticky, that is the model flattening the paper: re-roll on nano-banana-2 before touching the prompt. Never mix image models within one ad. - Video model:
advibly_generate_videowithmodel: "gemini-omni-flash". Omni Flash takes 4 to 10s durations (the 4 to 6s shot cadence fits), keeps baked text the most stable, and produces the best layered parallax on flat art. It is 9:16/16:9 only, has no end frame, and needs positive-only phrasing. Fall back toseedance-2.0(mode: "fast", any 4 to 15s, every aspect ratio) when you need an exact reveal landing viaend_image_url, a 1:1 or other non-vertical/horizontal aspect, or a shot past 10s. For real people or third-party brand logos switch the ad tokling-v3(3 to 15s, native audio): Omni Flash and Seedance block recognizable celebrities and marks at the content-filter level. One video model per delivered ad. - Audio in clips is SFX only, no spoken words. Paper foley (tear, rustle, whoosh, settle thunk) and quiet room tone. Explicitly forbid narration, dialogue, and lyrics in every motion prompt: the voiceover is mixed on top in Phase 7 and anything spoken in a clip collides with it.
- Voiceover and music generate in-platform.
advibly_generate_voiceovernarrates the script (xAI TTS: voices eve / ara / rex / sal / leo, 20+ languages, ~0.03 credits per 1000 characters) andadvibly_generate_musiccomposes the instrumental bed (MiniMax Music 2.6, 0.3 credits per track). Phase 7 assembles VO, a static music bed, and quiet clip SFX in one freeadvibly_render_compositioncall. - Faithful theme, not brand recolor: pass
on_brand: falseon every generation. The theme's palette is the whole point; the brand-kit board would recolor it.brand_idis still required on every call, and the run'sproject_idgoes on every call too (together they file the work as one project tile in the user's library); withon_brandfalse neither styles the output. The brand shows through the real product photo, the headline copy, and one accent color, never through a stamped logo. - No text-overlay tool. Every headline is baked in at image generation. If a headline
degrades, re-roll with
num_images(up to 4), shorten the words, or tryseedream-5-profor that one poster. Never plan to overlay text afterward. - Format: ask 9:16 vertical or 16:9 horizontal up front (1:1 is possible on Seedance and Kling if asked). Hold one aspect across every shot.
- Prompts cap at 2000 characters on both generation tools. The style block plus scene must fit; trim vocabulary, not the layering description.
- Tools are deferred. Load the exact Advibly tool schemas with tool search before the first call each session (search "advibly generate image", "advibly generate video", "advibly generate voiceover", "advibly generate music", "advibly stitch videos", "advibly get products", "advibly upload asset", "advibly add subtitles", "advibly create project", "advibly update project"). Confirm parameter names and supported durations against what loads rather than assuming.
The Advibly asset workflow (memorize)
An image from advibly_generate_image returns a public url you can pass straight into
start_image_url or reference_image_urls on the next call. No re-upload. If a call returns
status: pending, the media still renders in chat; call advibly_get_generation
(wait: true) only when you need the finished URL downstream (for keyframes feeding motion,
you always do).
To bring in a file the user owns: advibly_upload_asset with source_url (public link) or
data_base64 (small local files), plus brand_id. Returns a reusable url.
The one reference you carry through the whole ad is the product photo:
- Store brands (
brand_type: "ecom_store"):advibly_get_products, pick the product with the user, note its image URL. Other brands: a product photo fromadvibly_get_assetsor an upload. Pass this URL inreference_image_urlson every poster that shows the product and describe it as "the real product photo, unchanged and photoreal, composited into the collage". Do NOT passproduct_idto the generation tools: on an image call it forces edit mode against the raw photo and fights the collage; on a video call it replaces your start frame. - Pre-reveal and no-product beats get no product reference. A hook poster or problem poster that should not show the product yet must not carry the photo reference, or the model leaks the product in early.
Logo: keep the corners clean by default. Do not stamp a corner logo tile, do not pass the
logo into reference_image_urls (it makes the model invent a watermark). Cohesion comes from
the shared theme: same style block, palette arc, type treatment. If the user explicitly wants
their mark, place it on the beat(s) they name, usually the close, never all of them.
PHASE 1: INTAKE
One message, only what you still need:
- Brand:
advibly_list_brands. One brand: use it. Several: ask. None: send the user to advibly.com/onboarding (this skill reads a brand, it cannot create one). - Product to feature and its photo URL (asset workflow above). A pure brand-story explainer with no hero product is fine; skip the photo and the product-reveal beat.
- Format: 9:16 or 16:9 (ask; no default).
- The one claim: what does this ad argue ("not all creatine is equal", "your invoices should chase themselves"). No angle? Propose 2 or 3 from the brand brief and let them pick.
- Length: default ~32s (6 to 8 shots). 60s is the ceiling: the stitcher takes at most 12 clips, which is exactly 12 shots at ~5s.
Once the brand is resolved, create the run's project with advibly_create_project (brand_id
plus a deliverable-shaped name like "Acme Vox explainer") and pass the returned project_id on
every generate call of the pipeline (bake-off posters, keyframes, clips, voiceover, music, the
final composition) so the run lands as one tile in the user's library. If the user is continuing
an earlier run, find its project with advibly_list_projects instead of creating a duplicate.
PHASE 2: BEAT MAP (the one mandatory approval gate)
Read references/beat-layer.md first. Pick the narrative arc that fits the claim
(pas/bab/aida for ads, how_it_works for product explainers, timeline for history,
myth_buster to correct a belief, man_in_hole for transformations). Then draft the full
beat map and show it for approval before generating anything.
Rules that make the map good (details and vocab in the reference):
- Beat 1 is a hook that lands in under 3 seconds, headline carrying the payoff promise.
Never spend beat 1 on setup. Pick a hook pattern (
surprising_stat,mistake_callout, ...). - Each beat gets 2 shots: a wide establishing poster carrying the headline, then a detail cut-in without it. The narration line spans both shots; the visual cuts mid-sentence. Shots run 4 to 6s, never past 7.
- Product ads place the product-reveal beat where the arc turns (the Solve in
pas, the Bridge inbab): the real photo arrives framed by a burst, converging arrows, or a glow. - Camera moves are varied: no two adjacent beats share a move,
staticis reserved for the payoff beat.element_motionis written fresh per shot to fit that scene, rich, several things moving. - Narration lines together are the ad's script. Write them as one continuous persuasive read, ~2.5 to 3 words per second inside each beat's total duration.
Deliver the map as JSON the user can edit field by field:
{
"project": "acme-creatine-vox",
"topic": "not all creatine is equal",
"brand_id": "<id>", "product": "Acme Creatine", "product_photo_url": "<url>",
"aspect": "9:16", "language": "en",
"arc": "pas", "theme": "<set in Phase 3>", "image_model": "<set in Phase 3, nano-banana-2 default>",
"video_model": "gemini-omni-flash", "motion_style": "punchy", "constraints": "strict",
"voice": "leo", "music": "driving cinematic percussion build, warm resolve, modern ad underscore",
"beats": [
{
"id": 1, "role": "hook", "title": "NOT ALL EQUAL",
"bg": "bold red", "feel": "urgent, confrontational", "hook": "mistake_callout",
"product": false,
"narration": "Most creatine on that shelf is not what the label says it is.",
"shots": [
{"id": "a", "dur": 5, "title": true, "shot_size": "WIDE", "camera_move": "push_in",
"scene": "a wall of near-identical grey paper supplement pouches, one torn-open gap in the grid, question-mark scraps",
"element_motion": "the grid of pouches shivers in place, one pouch peels off the wall and tumbles, halftone dots pulse"},
{"id": "b", "dur": 4, "title": false, "shot_size": "CLOSE", "camera_move": "parallax",
"scene": "close cut-in of one grey pouch with a magnifying glass cut-out over its label, torn newspaper scraps behind",
"element_motion": "the magnifying glass slides across the label, scraps drift at different depths"}
]
}
]
}
Per-beat product: true marks which posters composite the real photo. voice picks the
narrator (eve: energetic / ara: warm / rex: confident / sal: smooth / leo: authoritative)
and music describes the instrumental bed; both are generated in Phase 7. Ask: "Approve
this beat map, or edit any field?" Only proceed on a yes.
PHASE 3: THEME + MODEL (the user picks by eye)
Do not reuse one house style for every brand. Read references/prompt-guide.md §5 and pick
3 or 4 theme presets that fit the claim's era, culture, and tone (american-retro,
swiss-modern, punk-zine, soviet-constructivist, wpa-propaganda, 70s-groovy,
chinese-ink, atomic-age, paper-craft-cream), or compose a custom theme from the
dimension banks when none fit. Match the topic and brand, not the language.
Then run a bake-off: generate beat 1's wide poster once per candidate theme (3 or 4
advibly_generate_image calls, one image each, model: "nano-banana-2", on_brand: false)
and let the user pick by eye. AI proposes, the preset library is the quality floor, the human
decides.
The bake-off also settles the image model. nano-banana-2 is the default because it holds
the paper texture; if the winning poster's headline looks weak or the topic is type-heavy,
re-render that one theme on gpt-image-2 (or seedream-5-pro) beside it and let the user
compare texture against text. Lock the model that wins and use it for every poster (see
models-and-gotchas.md).
If the user wants to skip the spend, describe the candidates and let them pick from the
tables, but say the bake-off gives a much better read. Set the winner as "theme" (and the
chosen "image_model") in the beat map; its style block, palette, type, and motion amplitude
drive every prompt from here.
PHASE 4: KEYFRAMES (the collage look)
One poster per shot, in beat order. Compose every prompt with the 5-part structure in
references/prompt-guide.md §1: style block (identical on every poster), scene as separate
cut-out pieces, one bold flat background color, headline baked in on "title": true shots
only, aspect and resolution.
advibly_generate_image
prompt: <5-part collage prompt>
brand_id: <brand id>
project_id: <project id>
model: <the bake-off winner, "nano-banana-2" by default>
resolution: "2K" # add quality: "high" only if the model is gpt-image-2
on_brand: false
aspect_ratio: <9:16 or 16:9>
reference_image_urls: [<product photo URL>] # only on beats with "product": true
- Verify each poster is a real layered collage before animating: distinct pieces, visible torn edges, drop shadows, crisp headline. Re-roll here; images are cheap next to clips.
- Product posters: the photo reference plus the "unchanged and photoreal" wording. Check the product did not get redrawn as paper; that is the most common failure.
- The style block stays verbatim across every poster; only scene, background color, and headline change. That is what makes 12 shots feel like one film.
- Show the set, ask "keep or change?", re-roll only the misses.
PHASE 5: MOTION (living collage)
Animate each approved poster with the 5-axis motion prompt from
references/prompt-guide.md §2: goal, camera (the beat map's camera_move, one move only),
element movement (the beat map's element_motion, rich), aesthetic preservation, feel and
color, then the stability constraints.
advibly_generate_video
prompt: <5-axis motion prompt, SFX-only audio direction, no spoken words>
brand_id: <brand id>
project_id: <project id>
model: "gemini-omni-flash"
aspect_ratio: <9:16 or 16:9>
duration: <the shot's dur, 4 to 10>
start_image_url: <that shot's poster URL>
constraints: "strict"in the beat map means the defect guards stay in every prompt (flat 2D, one continuous move, no morph, text anchored)."loose"drops them for a bold move (orbit, dolly zoom, whip) on a beat that earns it; expect to re-roll.- Omni Flash takes positive-only wording: convert every "no X" into a positive ("the camera stays locked", "the lettering stays exactly as printed"). See §2 of the prompt guide.
- The dramatic looks are motion prompts, not a separate engine. Pieces assembling from
an empty field, confetti and paper scatter, an impact shake, a whip: all reachable by
pushing the element-motion line, phrased for the model. Recipes and the phrasing bank are
in
references/motion-collage.md. Use them on the beats that earn a punch (hook, product reveal, payoff, ending), never every shot. - Exact reveal landings and the hard assemble-from-empty: when a beat must end on a
precise revealed frame, or you want a true bare-field-to-finished-poster build, that needs
end_image_url, which only Seedance supports; run the whole ad onseedance-2.0(mode: "fast") in that case and keep one model across the ad. Seemotion-collage.md§1. - Audio direction in every prompt: paper foley only, no voiceover, no spoken words, no music with lyrics.
- Show each clip: keep, re-edit (same poster, adjusted motion prompt), or re-roll.
PHASE 6: VOICEOVER + MUSIC + FINAL COMPOSITION
- Voiceover. Join the beats' narration lines into one continuous read and generate it:
If the read exceeds the scene total, tighten and regenerate it; never time-stretch.advibly_generate_voiceover brand_id: <brand id> project_id: <project id> text: <the full narration, beats joined in order; use [pause] between beats when a beat's line lands short of its window> voice: <the beat map's voice: eve / ara / rex / sal / leo> language: <only when auto-detection would get it wrong> - Music. Generate the bed from the beat map's
musicdescription:
Tracks may run longer than the ad; composition auto-trims with a tail fade. Keep it instrumental: lyrics fight the narration.advibly_generate_music brand_id: <brand id> project_id: <project id> prompt: <the beat map's music description + tempo + "modern ad underscore"> instrumental: true - Compose once. Call
advibly_render_compositionwith ordered clips asscenes(eachvolume: 0.3),voiceovers: [{source: <VO generation id>}], the music generation asmusic, the chosenaspect_ratio, the run'sproject_id, andkeep_scene_audio: true. The default static music level already sits correctly under narration. It returnsstatus: pending,generation_id, andedit_url; let the chat widget poll the render. - Captions (optional, only after composition). Use
advibly_get_generationwithwait: trueonly here to obtain the finished render, then calladvibly_add_subtitleswith a dynamic preset (glide,fusion,glass). Never caption the SFX-only cut; there is nothing to transcribe. Add brand words tovocabularyso the transcriber spells them right. - Deliver the result and mention the
edit_urlso the user can fine-tune the ad in the Advibly video editor. - Set the project cover. Call
advibly_update_projectwith{ project_id, cover_generation_id: <the final composition's generation id> }so the project tile shows the finished ad.
PHASE 8: OPTIONAL PUBLISH
If the user wants to post it: advibly_social_list_accounts, then advibly_social_create_post
with the final video. Only offer after the user has seen the finished ad.
Notes and rules
- The beat map is approved before any generation. The story is the product; the collage renders it.
- Real product stays photoreal; everything else is printed paper. Pass the photo reference on product beats, forbid it on pre-reveal beats, and say "unchanged and photoreal" in words every time.
- Cut every 4 to 6 seconds. Two shots per beat, wide plus detail, narration spanning both. One long static poster per beat reads as dead air.
- No two adjacent beats share a camera move;
staticis the payoff. Anti-monotony is the biggest quality lever after the posters themselves. - Element motion is where the energy lives. Write it per shot, rich, several elements moving. A hero element flying across the frame is an occasional punch, not a formula.
- One image model and one video model per delivered ad. Swap the whole set or nothing. Omni Flash is the default motion model; move the ad to Seedance only for end-frame reveals, non-9:16/16:9 aspects, or shots past 10s, and to Kling for real people.
- SFX-only clips. The voiceover and music generate in Phase 7 and mix on top; never let a clip speak.
on_brand: falsealways;brand_idandproject_idalways; no logo watermark by default.- Self-contained prompts. Generators have no memory of earlier calls; the style block travels verbatim in every image prompt.
- On failure:
content_rejectedmeans the policy blocked the prompt; rework wording, and if the trigger is a real person or third-party mark, move the ad tokling-v3.insufficient_credits: calladvibly_buy_creditsand share the checkout link. A stubborn reveal: switch that beat to the Seedance start-plus-end-frame route. - No em dashes, minimal emoji in any copy, label, or narration you draft.
Reference files
references/beat-layer.md: the STORY layer. Narrative arc library, hook patterns, beat counts per duration, shot sizes, the flat-safe camera-move vocabulary, element motion, and the anti-monotony move-rhythm presets. Read before writing any beat map.references/prompt-guide.md: the LOOK layer. The 5-part image prompt structure, the dimension vocabulary banks, the 5-axis video prompt with stability anchors, the advanced motion vocabulary, and the theme presets with the bake-off procedure. Read before writing any prompt.references/motion-collage.md: the dramatic looks (pieces assembling from an empty field, hero fly-across, confetti and scatter, impact shake, whip) and the image-model background-remover recipe, all done through the MCP with no local scripts. Read before reaching for a punch bigger than living-poster motion.references/models-and-gotchas.md: image and video model choice, content blocks, the positive-phrasing rule, the asset workflow, and the final composition contract (a static low music bed, tail protection, VO-overrun fix, whip assembly). Read before debugging a weak render or the audio mix.
版本历史
- c3c0a1e 当前 2026-07-23 11:05
