# Viral-dance short — agent playbook ⭐ flagship > Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. **Pick a dance style, pick a cast of characters — get a 10-second TikTok-ready short.** Build a vertical 9:16 reel of original characters performing a viral-style synchronized dance under neon lights, with an energy-matched music beat and a TikTok-safe caption overlay. Ready to upload to TikTok / YouTube Shorts / Instagram Reels / Douyin. ## What you'll get A 10-second (default) social-share dance short, built end-to-end via the storyboard MCP: - **One ensemble hero keyframe** (gpt-image, 9:16 vertical) — N characters staged with VERTICAL DEPTH (back/middle/front layering), not horizontally side-by-side. This is the critical composition recipe — see Lesson 1 below. - **One animated dance clip** (seedance-i2v or kling-v3-i2v, 10 seconds) — synchronized choreography described in abstract movement primitives (sway / hop-step / arm-wave / spin) rather than copyrighted choreography. The agent rewrites any specific viral name into renderable equivalents. - **One TikTok-ready music bed** (`music`, minimax) — structured brief with explicit BPM + drop timing + genre, passed with `lyrics_prompt: "[Instrumental]"`. See Lesson 3 below. - **One bold caption overlay** (`hyperframes-caption`) — placed in TikTok's UPPER-MIDDLE SAFE ZONE (y≈300), NOT at the bottom (which TikTok's own UI covers). See Lesson 4 below. - **A showcase HTML page** with the reel embedded, the cast described, choreography breakdown, pipeline trace. **Time**: ~5–10 minutes wall clock (i2v queue varies). **Your attention**: ~3 minutes (one approval gate on cast + dance style before fire). **Cost**: ~$2–$3 per reel. *Fastest flagship in the catalogue* — single keyframe + single i2v + single music + single caption. --- ## Top-1% TikTok-quality lessons (HARD-WON, do not skip) ### Lesson 1: VERTICAL COMPOSITION RECIPE — stack characters by depth, never side-by-side **The trap**: when an agent renders N characters in a 9:16 keyframe, the default instinct is to place them in a horizontal row (left-center-right). This DESTROYS the reel. **Why**: ALL current i2v models (kling-v3-i2v, seedance-i2v, ltx-i2v, veo-i2v) normalize their output to ~4:3 (1108×828 or 1112×834) regardless of source aspect ratio. When ffmpeg then center-crops 4:3 back to 9:16 for TikTok, the LEFT and RIGHT thirds get sliced off — meaning only the center character survives. **The fix**: stage characters with VERTICAL DEPTH: - **Back / top** — tallest character standing in the back, head + upper body in the top third - **Middle / center** — full body of the centerpiece character, mid-third - **Front / bottom** — closest-to-camera character (smaller due to perspective), bottom third All three characters then live in the CENTER COLUMN of the keyframe, which is exactly what survives center-crop. Verified on MONSTER DANCE CREW after the first run failed. ### Lesson 2: i2v model selection — they ALL center-crop to ~4:3, so plan for that **Reality**: kling-v3-i2v, seedance-i2v, veo-i2v — they all output their own internal ratio (~4:3) regardless of 9:16 source. There is no "preserves 9:16 natively" i2v on BYOC today. **Implication**: stop fighting it. Compose with the assumption that the final reel is a 1080×1920 center-crop of a 1112×834 video. Anything important must live in the center column. Lesson 1 enforces this. **Model choice**: - **kling-v3-i2v**: fast (60–120s), reliable, decent motion. Default choice. - **seedance-i2v**: better character consistency for cartoon characters; slower (200–300s); occasional long queue waits. - For dance reels, kling is usually fine — the prompt-following is what matters, not the model. ### Lesson 3: Music brief MUST be structured, NOT vibe-only **The trap**: a brief like "energetic playful spooky-fun beat" sounds specific but produces generic kitsch. **The recipe**: every music brief must specify: 1. **BPM**: exact number (120 for shuffle/dance, 128 for trap/hyperpop, 100 for samba) 2. **Genre + reference**: "modern hyperpop / electro-trap", "K-pop point-dance trap", "Latin samba" 3. **Drop timing**: "intro 0–2s, drop at 2s, peak 2–8s, fill 8–9.5s, stop-drop 9.5s" 4. **Instrumentation**: "808 bass + finger-snap claps on off-beats + pluck-synth lead hook" 5. **Anti-guide**: "NOT cute, NOT kitschy, NOT spooky" — name the kitsch trap explicitly Without the BPM + drop-timing combo, the music doesn't match the dance rhythm and the reel feels disconnected. ### Lesson 4: TikTok caption placement — UPPER-MIDDLE, not bottom **The trap**: placing the caption at the bottom of the 9:16 frame (y≈1450–1700) is invisible after TikTok's own UI overlays its caption, share buttons, and music icon on top. **The TikTok safe zone**: - Top 0–130px → unsafe (TikTok status bar) - **Upper-middle 200–500px → SAFE — put captions here (y=300 default)** - Middle 500–1400px → safe but covers the main video - Bottom 1400–1920px → unsafe (TikTok own caption, share UI, music icon) **The recipe**: caption PNG with bold Impact / Helvetica-Bold ≈ 92pt, white with heavy black drop-shadow (8 directions × 4px offset for legibility on any background), centered horizontally, anchored at **y=300**. ### Lesson 5: Choreography vocabulary — abstract primitives, never named viral routines **The legal guard**: do not pass "do the Macarena", "Single Ladies bridge", named choreographer's signatures, etc. The playbook rewrites every specific viral name into generic verbs the AI can render: - sway side-to-side (hip-shift) - hop-step forward (with optional twist) - arm-wave (sequenced left-then-right) - spin / turn / twist - starfish pose / V-pose / final pose - shoulder-shimmy / chest-bump-with-self Combine 4–6 of these into a 10s choreography. The AI will render group-synchronized versions reliably. ### Lesson 6: NO IP characters, EVER The playbook explicitly refuses Pixar / Disney / Studio Ghibli / anime-franchise / named brand mascot likenesses. If a user asks for one, the agent offers to **invent an ORIGINAL crew** in the same spirit. The MONSTER DANCE CREW example came from a "do the Monsters Inc dance" brief — the agent refused, invented Bramble + Sprig + Ozzie, and built the reel from those. The result reads as a "buddy monster crew" without copying any specific franchise. --- ## Tell the agent about the cast + dance ```yaml characters: # 2-5 character specs. The agent renders each freshly per run. # Each spec ~30 words, with EXPLICIT depth-layer assignment: # - name: Bramble # body: tall lavender-purple fuzzy monster # face: three small round eyes, floppy ears # layer: back-top # back-top | middle-center | front-bottom # pose: arms raised in V-shape above head # # The agent enforces vertical-stacking composition (Lesson 1). dance_style: # Movement primitive list (Lesson 5). Examples: # - "synchronized neon-shuffle: sway side-to-side, then arm-wave, # then hop-step forward, then final V-pose" # - "K-pop point dance: synchronized finger-snap arm-cross, hip-shift, # then point-and-pose finale" # The agent REFUSES copyrighted choreography names and substitutes primitives. setting: # Visual context. Examples: # - "Neon-lit nighttime city alley, magenta + teal signs, full moon, ankle fog" # - "Sunset rooftop, golden hour, distant city skyline, hazy" # - "Arcade floor, CRT screens glowing, retro-pixel aesthetic" caption: # Single hook line, ≤6 words. Bold sans, ≈92pt, white + drop-shadow. # Placed at y=300 (Lesson 4 — UPPER-middle TikTok-safe zone). music_brief: # MUST follow Lesson 3 template — BPM + genre + drop timing + instrumentation # + anti-guide. Example for zombie-shuffle: # "120 BPM modern TikTok-shuffle hyperpop / electro-pop. Intro 0-2s with # finger-snap pickup, drop at 2s, peak 2-8s with 808 bass on every kick + # finger-snaps on off-beats + pluck-synth lead hook, fill 8-9.5s, # stop-drop 9.5s. NOT cute, NOT kitschy, NOT lo-fi. No vocals." duration_sec: # 10 (default) / 8 / 6. dance_slug: # kebab-case for output. e.g. monster-dance-crew / robot-shuffle / fox-line. ``` --- ## Quality controls (auto-applied) The pipeline guards inherited from THE EMISSARY / THE ROAD / 麻婆豆腐 family plus dance-specific additions: - **Brief-stage IP refusal** — Pixar/Disney/Ghibli/anime franchise characters refused; agent offers original crew. - **Brief-stage choreography abstraction** — named viral routines rewritten into generic verbs. - **Vertical-stacking composition** — keyframe prompt explicitly specifies back-top / middle / front-bottom layer assignment per character. - **Center-column character placement** — every character's body must be in the center 60% of width, so center-crop preserves the cast. - **Music brief template** — BPM + genre + drop timing + instrumentation + anti-guide required; vague briefs rejected. - **Caption TikTok-safe placement** — y=300 (NOT y=1450); upper-middle zone above the TikTok UI overlay area. - **Single shot** — camera stays still, no mid-clip cuts (i2v doesn't do cuts cleanly). - **Mid-render review gate** — agent surfaces the keyframe AFTER it lands, before firing the expensive i2v wave, so a wrong-composition keyframe can be redone for $0.20 instead of a wasted $1.10 i2v render. > **Why this playbook is the fastest flagship.** One keyframe + one i2v + one music + one caption. No multi-shot stitching, no per-scene voice-over, no episode arc. Just a single 10-second bang. The constraint set above is what keeps that single bang from looking janky. ## The pipeline (real MCP verbs — 5 calls) ``` 1 — KEYFRAME create_media(action:"generate", model_override:"gpt-image", prompt:) → hero keyframe. ⛔ Mid-render gate: surface it before firing i2v. 2 — ANIMATE create_media(action:"animate", model_override:"seedance-i2v", source_url:, resolution:"auto", duration:, prompt:) → dance clip (~4:3 internal — that's fine, Lesson 2). Fallback on failure: kling-v3-i2v, then ltx-q-i2v. Always resolution:"auto". 3 — MUSIC create_media(action:"music", model_override:"music", prompt:, lyrics_prompt:"[Instrumental]", duration:) → music.mp3. lyrics_prompt:"[Instrumental]" is REQUIRED or it sings your style brief. 4 — CAPTION create_media(model_override:"hyperframes-caption", source_url:, text:) # bold, y≈300 upper-middle (Lesson 4) → captioned clip. (Or render a caption.png and ffmpeg-overlay it at y=300.) 5 — FINISH a) create_media(action:"mux_audio", model_override:"ffmpeg-mux", source_url:, audio_url:, audio_fill:"loop") → MANDATORY: audio_fill:"loop" makes the beat cover the FULL 10s (a bare mux leaves a silent tail when the i2v clip outlasts the track). b) create_media(model_override:"ffmpeg-export", source_url:, params:{preset:"tiktok-portrait", crop:"center-crop"}) → 1080×1920 9:16 reel, center-crop preserving the center column (Lesson 1+2). ``` The whole thing is 5 `create_media` calls. Step 5a's `audio_fill:"loop"` and Step 5b's center-crop are the two finishing moves that separate a clean reel from a janky one. ## Output structure ``` public/chapters// keyframe.png ← hero shot of the cast (vertically-stacked composition) caption.png ← TikTok-style caption overlay PNG music.mp3 ← the dance-beat score -short.mp4 ← final 10s vertical reel public/chapters/-example.html ← showcase page ``` The example for this playbook is **MONSTER DANCE CREW** — an original three-monster crew (Bramble, Sprig, Ozzie) doing a synchronized neon-lit shuffle. The first pass was rebuilt after the user feedback exposed three bugs (music too kitsch, characters horizontal-staged so center-crop killed two of three, caption in TikTok UI dead-zone). The lessons above are baked from that rebuild. See `/chapters/monster-dance-crew-example.html`.