skills/krea-marketing/workflows/launch-teaser.md

# Launch Teaser

## Trigger

User says "launch video", "brand film", "teaser", "hype video", "product reveal", "kinetic type video", "make it cinematic", or "make it ultracool", or asks for a website-hero or YouTube pre-launch video. Disambiguate from `social-video-short.md`: that workflow is creator-native, single-continuous-clip social content; this one is a designed multi-beat film with typography and music. A clean multi-shot product commercial with NO on-screen type and no music composition - especially one replicating a reference ad's cut structure - routes to `cinematic-product-ad.md` instead. When in doubt: if the video needs on-screen type, beat-synced cuts, or "premium/agency" energy, route here.

## Clarify

Ask the user once, in a single batched message. Skip whichever the user already volunteered.

- **Aspect + placement**: 16:9 website hero/YouTube, 9:16 Reels, or 16:9 master plus vertical cut.
- **Duration**: ~15s teaser (default) or ~30s film.
- **Footage**: generate fresh cinematic shots (default), reuse existing campaign stills with motion treatment, or typography-only.
- **Sound**: music-driven (default), music plus VO, or silent-first.
- **Brand name + tagline** for the end card. Ask explicitly - a launch film wants a name reveal; if there is none, stay product-led.

If the user gave a tight, complete brief, skip Clarify entirely and proceed to Recipe.

## Recipe

Hard prescription. Follow in order - the order IS the deliverable. The quality comes from separating AI footage from designed motion: video models generate raw, textless, cinematic material; ALL typography, cuts, flashes, HUD overlays, and beat sync are built deterministically in an HTML composition layer on top. That layer is the `hyperframes` skill; if it is not installed, install it first: `npx skills add heygen-com/hyperframes --skill hyperframes`.

1. **Lock the beat grid first.** Pick a rhythm template from the `hyperframes` skill's beat-direction patterns before generating anything; for 10-20s teasers default to SLAM-proof-SLAM-hold. Write down exact cut timestamps (for 15s, e.g. 3.0 / 6.2 / 7.4 / 10.8 / 13.0). Everything downstream serves these numbers.
2. **Pick a visual style** from the `hyperframes` skill's visual-styles presets - Shadow Cut for dark/dramatic, Maximalist Type for loud hype. The product's own aesthetic decides.
3. **Generate music against the beat grid** with the ElevenLabs Music API (`music` skill). Write the cut timestamps INTO the prompt as explicit musical events: "hard drop hit at exactly 3 seconds", "final massive impact at 13 seconds that rings out until the end". Without explicit timestamped hits the API returns rhythmless wallpaper. Verify duration with ffprobe - the API may return ~1s extra; the composition's `data-duration` crops it.
   - **ElevenLabs tier fallback**: free-tier keys get 402 `paid_plan_required` on `/v1/music`, but `/v1/sound-generation` usually works - a degraded fallback for short atmospheric beds. Validate keys with GET `/v1/user`; a 401 `missing_permissions` for `user_read` still allows generation endpoints, so test the actual endpoint before declaring a key dead.
4. **Build start-frame stills, image-to-image from the product reference** - never text-only; product fidelity dies. Resolve the still model from the marketing image set in `../SKILL.md`. One still per footage beat; 3 shots is plenty for 15s. Prompt scene, lighting, camera, motion feeling, and composition only; do not include product material, color, silhouette, trim, hardware, label, or garment descriptors in the generation prompt. Hard prompt rules: "no text, no logos" in every still and every video prompt. Cinematic vocabulary is ENCOURAGED here - this is the opposite of the UGC workflow's banned-words list; do not cross-contaminate.
5. **Blocking still QA gate**: inspect each start-frame still with vision before animation. Reject product drift, Prompt-text override, generic product substitution, or any generated text/logo artifact. Do not submit video jobs until the stills pass.
6. **Animate each still into a 4-5s clip** via the start_image path on a cinematic video model resolved from live `list_models`. One steady camera move per shot - push-in, orbit, or pull-back; the composition handles all dynamism. Pacing guardrail still applies: never `slow`, `gentle`, `soft` - models literalize them into slo-mo; use `natural realtime`, `smooth`, `steady`. Footage only needs to be slightly longer than its slot; trimming is free in composition.

   **Real packaging/label text survives animation when the shot is built right.** "No text, no logos" (step 4, above) bans asking the model to invent new typography - it does not mean real printed text already on the product has to be avoided. Verified: a product with dense printed packaging text, animated from a high-fidelity still with a pure camera-only move ("nothing else changes, text and label stay exactly as shown") stayed fully legible through the entire clip, including the last frame. The failure mode isn't "video can't hold text" - it's motion that changes the scene (contact with liquid, melting, added elements) letting the model drift. Keep motion to camera-only for any shot where label legibility matters, and inspect the LAST frame specifically, not just the first - drift is progressive.
7. **Compose in HyperFrames.** Route to the `hyperframes` skill for all authoring rules; do not improvise composition syntax here. The composition must contain:
   - type scenes between/over footage on the locked beat timestamps;
   - punch-in cuts: each footage clip enters at 1.15-1.25x scale, snaps to 1.0 on the beat, then slow-creeps;
   - flash frames (1 white default; at most one colored accent, used once) and optional glitch/RGB-split bursts, HUD brackets, or scanlines per the chosen visual style;
   - end card: name/tagline reveal on the final hit, fade to black.
   - **Known trap**: video elements must NOT be nested inside timed divs - that violates the frozen-video contract. Wrap each video in a NON-timed div for scale/punch animations and put `data-start` on the video itself.
8. **Gate the render**: lint -> validate -> inspect -> render with HyperFrames. No skipped gates.
9. **Inspect rendered frames**: extract 4-6 frames with ffmpeg at the beat timestamps; check type legibility, clipping, safe areas, and product fidelity. Fix and re-render once.
10. **Deliver** with a beat-map summary - what happens at each timestamp. Offer the vertical cut.

### Tool path

Inspect live model schemas before submitting, and use only fields exposed by the available Krea tools. The `hyperframes` skill owns all composition details. Skeleton:

```text
1. Upload the product reference to Krea; use the returned Krea-hosted URL.
2. Batch start-frame stills: marketing image model, image-to-image from the
   reference, one per footage beat, scene-only product prompt, "no text, no logos"
   in every prompt.
3. Inspect stills and reject product drift before video.
4. One video job per shot: cinematic video model, start_image=<still>,
   4-5s, one steady camera move, natural realtime.
5. curl -X POST https://api.elevenlabs.io/v1/music with the cut timestamps
   written into the prompt as explicit hits; ffprobe the returned file.
6. Compose, then run the HyperFrames lint, validate, inspect, and render gates.
```

## Banned

- Do not ask the video model to render typography, logos, UI text, or exact timing - it does so unreliably; composited type is pixel-perfect on every render.
- Do not generate music before cut timestamps are locked.
- Do not generate stills text-only when a product reference exists.
- Do not animate stills that have not passed product-fidelity inspection.
- Do not let UGC realism vocabulary (handheld, amateur, phone-framing) leak into this workflow.
- Do not skip the composition layer and deliver raw concatenated AI clips as a "launch film".

## Cost & time

- Per-job: ~3 stills + ~3 short video clips + 1 music track + 1 composition render.
- Each 4-5s video-model shot is roughly half the credit cost of one 15s UGC clip.
- Wall-clock is dominated by footage (~10 minutes) and the composition render (2-8 minutes).

## On failure

| Symptom | Cause | Fix |
|---|---|---|
| Music has no hits, rhythmless wallpaper | Prompt lacked explicit timestamped events | Re-prompt with harder timestamp language: name every cut, "hard drop hit at exactly Ns" |
| 402 `paid_plan_required` on `/v1/music` | Free-tier ElevenLabs key | Fall back to `/v1/sound-generation` for a short atmospheric bed, or ask for a paid key |
| Video frozen in render | Video element nested inside a timed div | Wrap the video in a non-timed div for punch animations; put `data-start` on the video itself |
| Type clipped or illegible | Safe areas violated | Safe-area margins >= 100px; captions >= 26px at 1080p; re-inspect before re-render |
| Product drifts in stills | Weak reference anchoring or prompt-text override | Remove product descriptors from the prompt, keep the reference image primary, and retake before video |
| Recolored logo | Product still flattened into a tinted graphic | Reject still and retake with photo-first scene prompt |
| Animated clip invents an extra part on the product (a charm, a drip, a redesigned detail) | `start_image` alone does not anchor identity for the full clip duration; on longer atmospheric motion (water, reflections, moody lighting) the model drifts small details toward whatever aesthetic the scene suggests, even with the product held at a safe distance | Pass the original product reference in `reference_images` alongside `start_image`, and repeat exact-preservation language in the motion prompt itself ("keep the [product detail] exactly as shown, do not redesign it") — not just in the still-generation prompt. Verified fix: this alone resolved a pendant that had drifted into an invented design. Re-inspect the last frame specifically, not just the first — drift is progressive. |
| Video job fails with `content_policy` on a still that already passed inspection | Dark/moody single-source rim light plus an orbiting camera has hit provider-side moderation false positives | Retry once with simplified, brighter wording; if it fails twice, swap to a brighter treatment instead of retrying indefinitely |
| Footage is slow motion | Banned pacing words in prompt | Rewrite with `natural realtime`, `smooth`, `steady` |