gooseworks-ai/goose-skills / skills/video/video-production-harness
video-production-harness
VerifiedFull video production sequence with script review, actual ingredient choices, controlled generation, editing, evidence-based quality review, polish, captions and delivery. A host binding supplies project storage, authentic approvals, provider access and billing.
references/shot-routing.md
View on GitHub# Human version
Pick every shot's model, or its kit part on the kit path, from one table. Re-roll the cheapest layer that fixes the problem.
---
# Agent version
The two paths are defined in capabilities.md at the package root. Agent-built path (the default): use the fal model id column. Kit path only: the part column. Rows with no model, agent-built: `create-kinetic-typography` for type, `goose-graphics` for cards and stills, FFmpeg overlay for screenshots and logos.
The **Known tells** and **Check before keeping** columns name what to look for on each row; grade each problem by [check severity](check-severity.md) (extra hands, a redrawn product, garbled text and a wrong mouth are `must_fix`; waxy skin or drift are `ask`), and re-roll only for a `must_fix` or a fix the customer (or you, in `review_mode: auto`) chose. Use this table in three places: the design brief writes one routing line per scene (`scene N → <row> → <fal model id>`, plus `part@version` on the kit path), create-clips makes each scene on its row, and edit-clip re-rolls on the same row. Then open the genre card for the brief:
- AI creator or UGC talking video: [ai-creator-ugc](genres/ai-creator-ugc.md)
- Kinetic type or designed motion cards: [motion-graphics](genres/motion-graphics.md)
- Fast product hero motion: [hypermotion](genres/hypermotion.md)
- Chat or app screen story: [chat-ui](genres/chat-ui.md)
- Product demo or proof points: [product-demo](genres/product-demo.md)
- A narrator (voice or cartoon cutout) over a montage of people using the product: [narrator-montage](genres/narrator-montage.md). Any reference with a narrator over people uses this card, not product-demo or ai-creator-ugc.
Before choosing a native-speech, lip-sync or narration row, read [speech and route qualification](speech-and-route-qualification.md); for product and person shots, read [product scale and world](product-scale-and-world.md). A row is a candidate: inspect the output. A shot with no row is an open decision for the operator, not a free model choice. For any shot with a person, also follow the avatar guide that ships with `create-image-fal`.
**Speaking route.** Read [talking-actor production](talking-actor.md). Exact-script preserves audio; single-native-take uses the kit. A voice-reference route needs its own quote/approval, never Fabric pricing or unsupported kit inputs. Only an on-screen speaking face needs audio-driven lip-sync; native speech needs no added pass.
**Hands and capture (people rows).** A shot whose action uses both hands (applying, smoothing, pressing, holding a product with both hands) is captured `phone_counter` (a propped phone or a mirror) or `phone_observer`, never `phone_selfie`, which makes the model add a hand holding the phone. On the creator-h3 route, use `phone_counter` (it has no `phone_observer`). Every person clip gets the hand count in [review-video](../review-video.md).
**Quality tier.** The customer's Standard or Best (`quality_tier`) sets each row's starting setting in the **Standard / Best** column. Read [quality tier](quality-tier.md) before choosing a setting, upgrading a shot or sending Seedance anything: it has the defaults (no video above 720p/768p unless the customer asks), the picture-only upgrade rule after a failed check and the no-face rule.
| Shot type | Kit part@version | fal model id | Prompt rules | Known tells | Cost | Standard / Best | Check before keeping |
|---|---|---|---|---|---|---|---|
| Talking creator, exact script | `voice-elevenlabs@1.3.0` → `lipsync-veed@1.1.0` | `eleven_v4`, `veed/fabric-1.0` | **For an exact supplied audio track across clips** (`exact_audio_lipsync`), only on shots marked `speaking_on_camera` ([lip-sync only for a speaking face](speech-and-route-qualification.md)): make and accept the final voice line first (rate-checked, approved) in the person's `character.voice_pair` voice, then one lip-sync per line, one line per clip, never re-synced after a picture edit; neutral closed-mouth still from the cast; the tier's resolution (480p Standard, 720p Best); both hands busy: never `phone_selfie` | Mouth smear on hard consonants; a line split across two clips | $$ | Fabric 480p / 720p | Each clip within 0.5 s of its line; listen to every line before lip-sync|
| Talking creator, single native take | `creator-h3@1.8.0` | `minimax/h3-max/reference-to-video` | Native H3 voice (`qualified_native`) only for one take of 15 s or less, at most 3.2 words/s; cut only between sentences; never retime the take; no square brackets in lines; pass `reference_words_per_second`; no voice pair (the person is made with `--no-voice-pair`, and the take's own voice is accepted with the first take); both hands busy: never `phone_selfie` | Rushed or clipped last words; a join mid-sentence; face drifts from the still | $$$ | not in a tier quote (it prices talking shots as Fabric): ask the owner first | Listen to the take (or measure it: word timing against the line); transcript matches the script; longer voice-reference work follows the separately qualified route above|
| Narrated creator scene montage | `video-h3max@1.3.0` (stills mode) | `minimax/h3-max/reference-to-video` | One still per scene is the first frame; write motion only; clips 4-15 s timed to the voice; drop native sound when the voice comes from another step; both hands busy: never `phone_selfie` | Animated black bars; product redrawn | $$$ | H3 768P / 768P (1080P only when the customer asks) | Each clip covers its scene; product matches its photo|
| Two-person dialogue / street interview | `video-h3max@1.3.0` (take mode) | `minimax/h3-max/reference-to-video` | Native dialogue with a speaker per beat; at most 4 takes of 8 beats; a spoken beat needs words ÷ pace + 0.5 s. Never send a photoreal person still to Seedance; both hands busy: never `phone_selfie` | A line from the wrong mouth; drift at take ends | $$$ | not in a tier quote (it prices talking shots as Fabric): ask the owner first | Each line comes from the right speaker; transcript matches|
| Seated two-person talk | `cast-plate@1.1.0` → `lipsync-veed@1.1.0` → `dialogue-cut@1.0.1` | `openai/gpt-image-2/edit`, `veed/fabric-1.0` | One plate for the room (one still per person puts them in different rooms); faces neutral, mouths closed; 2-3 people with different looks; both hands busy: never `phone_selfie` | Twins; mouth smear | $$ | plate medium / high; Fabric 480p / 720p | Singles differ; each clip within 0.5 s of its audio|
| Creator B-roll, no speech | `video-h3max@1.3.0` (stills mode, `broll: true`) | `minimax/h3-max/reference-to-video` | The still is the first frame; motion only; one `extra.action` per scene, a physical action that changes the picture (apply, peel, pour, stir), never a head-only one (looks, blinks, turns her head) or "hands still": the part refuses those and `generate_audio: true`; "nobody speaks" always comes with the named action; one camera move; 4-15 s; both hands busy: never `phone_selfie` | Lips moving with no line; hands morphing; content the prompt never asked for | $$$ | H3 768P / 768P (1080P only when the customer asks) | Watch the whole clip against its still; the named action happens; crop the face 2× at 25/50/75% for added gloss, halos or a light change; mouth closed|
| Lipsync onto an existing take or still | `lipsync-veed@1.1.0` | `veed/fabric-1.0` | Only a face on screen speaking that line (`speaking_on_camera`); never a silent face or voiceover over B-roll (use Creator B-roll); after the final line, once per line; neutral, closed-mouth, head-and-shoulders still; one approved audio line per clip; the tier's resolution (480p Standard, 720p Best); both hands busy: never `phone_selfie` | Mouth smear on hard consonants | $$ | Fabric 480p / 720p | Length within 0.5 s of the audio; watch the mouth|
| Cast anchor portrait | `image-gpt-image-2@1.3.0` (one person); `cast-plate@1.1.0` (2-3 in one room) | `openai/gpt-image-2`, `openai/gpt-image-2/edit` | Real people use the edit route; people come from create-creator-takes-h3's make_character.py (`--shot talking` for a talking head, `--action "<task>"` or `--capture phone_observer` for anyone who acts, `--skin-condition` when the product treats one). Capture, focus, light and texture from the approved look; use the phone realism block only for phone footage; frontal gaze; state each person's ethnicity in group shots; both hands busy: never `phone_selfie` | Waxy skin; blurred background; face under 300 px | $ | medium / high | Crop the face, enlarge 2× and inspect|
| Set / world reference still | `image-gpt-image-2@1.3.0` | `openai/gpt-image-2` | Empty set, no people; even light, deep focus; the reference for every shot there | The room changes between scenes | $ | medium / high | Compare each scene's background with it |
| Presenter or UI still | `image-gpt-image-2@1.3.0` + `overlay@1.0.0` | `openai/gpt-image-2/edit` | Leave screens blank or turned away; put the real screenshot on with overlay; no words in the prompt; both hands busy: never `phone_selfie` | Garbled labels; a border or letterbox | $ | medium / high | Read every string on the frame|
| Photoreal keyframe / product hero | `image-gpt-image-2@1.3.0` + `overlay@1.0.0` | `openai/gpt-image-2/edit` | Real products use the edit route: send the real product photo; describe the action and set, not the product; one colourway. A still packshot's label comes from the real photo through `overlay`; a pack in hand uses the row below; both hands busy: never `phone_selfie` | Product redrawn (logo, shape, colour); extra hands | $ | medium / high | Product matches its photo at 2×|
| Product in hand or in use (still) | `image-gpt-image-2@1.3.0` | `openai/gpt-image-2/edit` | The real product photo is reference image 1, front label facing the camera (agent-built: create-image-fal `--brand-product held --product-image <real photo>`; a static packshot is `--brand-product static`); each use state (open pack, contents, cooked dish) from its own real brand photo; never type label text, never ask for a blank, unprinted or back-facing pack; both hands busy: never `phone_selfie` | Blank or sticker-flat pack; contents unlike the real ones | $ | medium / high | Label and contents match the photos at 2× |
| Product photo overlay | `overlay@1.0.0` | — | Real product photo or logo (PNG, JPEG, WebP) over the scenes that name it; up to 8 items | SVG logo skipped; covers a face | free | free | Watch every window |
| Cartoon keyframe | `image-nano-banana@2.0.0` | `fal-ai/nano-banana`, `fal-ai/nano-banana/edit` | **Cartoons only**, never real people or real products; send `subject: cartoon` (the part refuses anything else and any real product photo); one style anchor as reference; the reference's brightness beats a hex code; no text | Colours lighter or darker than locked; garbled text | $ | not in a tier quote (it prices stills as GPT Image 2): ask the owner first | Compare with the style anchor |
| Product in hand / motion from a keyframe | `video-veo-3@1.0.2` | `fal-ai/veo3.1/fast/image-to-video` | **Warning (15 Oct 2026): fal retires this endpoint on 17 Nov 2026**; keep it until a bake-off picks a replacement; for product-only motion prefer `video-seedance-2@1.3.0` (no faces). Start frame from the product still; one governing camera idea; both hands already on the object, no tap or twist; say "stationary" for still scenes; both hands busy: never `phone_selfie` | Dark plates fail; floating object; background moves | $$ | refused on a tiered video (`custom_video_tier_setting`): use H3 768P (`video-h3max@1.3.0`) | Sample 8 fps; last second shows the right product|
| Complex product motion | `video-seedance-2@1.3.0` | `bytedance/seedance-2.0/image-to-video` | Product photo as start frame, no face in frame (Seedance refuses photoreal faces, whatever model made them); 720p (1080p only when the customer asks); say whether the camera or the subject moves; 4-15 s | Camera and subject motion swapped; morph between impossible angles | $$$ | H3 768P (`video-h3max@1.3.0`); Seedance 720p, no face, only with the customer's new yes / the same | Watch for morphs; use an FFmpeg cut when no real camera path exists |
| Hypermotion product | `video-seedance-2@1.3.0` (reference mode) + `html-frames@1.3.1` cards | `bytedance/seedance-2.0/reference-to-video` | Product photos as references; no person references and no face in any reference (Seedance refuses photoreal faces, whatever model made them); 720p (1080p only when the customer asks); fast hero moves; all words on HTML cards | Product redrawn; text in the clip | $$$ | H3 768P (`video-h3max@1.3.0`); Seedance 720p, no face, only with the customer's new yes / the same | Product matches its photo; no AI text |
| UI / screen insert | `cut-footage@1.2.1` or `overlay@1.0.0` | — | Real screen recording or screenshot; cut the window that shows the feature being said | Wrong screen for the line; a flash cut | free | free | The feature is visible while it is said |
| Chat screen | `phone-chat@1.3.0` | — | One message per scene as "Name: text"; skins: iMessage, ChatGPT, notes, notification stack | Text past the screen; typing out of sync | free | free | Every message readable at phone size |
| Kinetic type / motion card | `html-frames@1.3.1` (seed frames only) + `overlay@1.0.0` for real screenshots | — | Words mirror the voice; each card holds for its reading time | Words that differ from the voice | free | free | Read every card at phone size |
| End card | `end-card@1.1.1` | — | Real logo, never redrawn; up to 3 rows of 60 characters; hold at least 1.5 s | Logo stretched; CTA missing | free | free | Readable at phone size for 1.5 s |
| Voiceover | `voice-elevenlabs@1.3.0` | `eleven_v4` | Approved voice only; brand pronunciations; one line per scene | Mispronounced brand words; slower than the reference | $ | same | Listen to every line; `scripts/reference_rate.py --against` the reference passes |
| Music | `music-elevenlabs@1.1.0` | `music_v2_5` | Request `music_v2_5` explicitly; mood from the plan; ask for a continuous bed with no quiet sections | A silent gap mid-bed | $ | same | Measure the level over the whole bed |
| Captions | `captions-layer@1.5.0` | `fal-ai/whisper` | Word timings from the voice step; keep captions off the product | Captions across a cut or over the product | $ | same | Read each caption against the audio |
**Cost ladder.** Re-roll the cheapest suitable layer: local text/FFmpeg (free) → still ($) → image-to-video ($$) → premium physics, talking or native audio ($$$). Scene words use overlays; real static ads follow [specialist handoffs](specialist-handoff.md). Every camera move serves the ad's governing idea.
**Models seen in older runs with no part.** Move them to a row: `fal-ai/flux/dev` → `openai/gpt-image-2`; Seedance text-to-video with a separate lipsync → `creator-h3@1.8.0` (one creator) or `video-h3max@1.3.0` (two speakers); `fal-ai/nano-banana-pro` → `fal-ai/nano-banana` for cartoons, `openai/gpt-image-2/edit` for real people and products; Kling 2.5 Turbo is withdrawn → `video-h3max@1.3.0` stills mode for creator B-roll, `video-seedance-2@1.3.0` for product-only motion; Veo 3.1 Lite → `fal-ai/veo3.1/fast/image-to-video` (retiring 17 Nov 2026); `music_v1` → `music_v2_5`.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |