Scenario exposes a large video catalog through one MCP loop: text-to-video and image-to-video generators plus video-to-video editors, lipsync, upscalers, and deterministic cut/split/concat tools. Per-family contracts: , , , , , , , , , .

Connection and the core generation loop: see the skill. If a sibling skill named here is missing from your available skills, ask the user to install it (); unattended, proceed from tool schemas and flag the gap.

The source image already fixes the look, so prompt only motion, camera, and timing ("orbit left", "hold on the final pose"), changing one clause per retry. A first-frame field pins how the clip opens (a last-frame field, where offered, how it lands, and only beside a first frame: alone it is a 400 ); a reference-image field carries identity and look without fixing any moment. So an exact opening composition goes in as a rendered still and a recurring character or product as references; whether one run takes both is per schema, and several forbid the pair. A reference-audio field is guidance the prompt has to claim: what it conditions (timing and energy on one family, a voice on another) and whether it rides alone or needs an image or video reference beside it are the family skill's contract, but on every member a track passed silently beside a character reference is easily ignored, so the prompt says what the track is for ("the reference audio is her own voice, lip-synced" for speech; the beat the cuts land on for music). Trained LoRAs do not ride along: at authoring time no video member took , so a custom style reaches video through a still generated with the LoRA and passed as the first frame or a reference. Iterate at the cheapest resolution tier and shortest useful , then upscale the keeper: a re-run at a higher tier is a new generation, since many video schemas expose no and none promises one reproduces a shot across tiers.

Write the prompt as a shot direction in sentences, not tags: shot size, setting, subject and one action, one named slow camera move, light source, style, and an audio line, in the order the family skill gives. Named moves and concrete light survive generation; mood words do not.

Many generators return sound with the picture (dialogue, effects, ambience, sometimes a score), switched by or where the schema has one, each with its own default; with none, the search hit says whether sound comes back, and the prompt is the only lever over what does. Effects and ambience go in plain words, dialogue in quotes, plus a generic "no music", never a named genre or instrument (the score is laid once in assembly; switch the audio off when assembly supplies the whole soundtrack) and "no subtitles, no on-screen text", since a caption tool adds captions and removes none. Exclusions ride the prompt unless the schema has . Type the viewer must read is composited in assembly (): generated type drifts.

All editing is on a video-input model found with :

Dubbing translates the speech and keeps each speaker's own voice, tone, and timing. It does not move the mouth, so a dubbed talking head still has lips forming the original language. Three steps, after any trim the limits below require:

Judge a localized talking head on a stylized character before promising it on a photoreal one.

Where a model bounds input length it rejects rather than trims: a 30.08 second reference against a 30.0 second limit fails the whole run, with the error naming both numbers. A clean prices the payload and is not proof it clears a ceiling: read the input's duration off and compare it to the cap yourself. A ceiling can be a typed on the file field, prose in that field's description, or absent, so check both, on the audio input as readily as the video. When one applies, trim with the deterministic cut or split tools (, , fixed ids for the same reason as the extractor above) before the run that enforces it, and before any step whose output must match the trimmed footage, such as a dub. Land inside the stated range, not on its edge.

Floors bind as hard as caps. The shortest clip a member renders is in its schema, as a on a numeric , as the smallest of a enum's , or as a fixed length when the field is absent (5 seconds on one reference-conditioned member, 1 or 2 on others at authoring time), so a 1 to 2 second VFX element is not a request on the first pick: with the clip length in the user's words, read the floor off the schema, and when nothing goes low enough render the shortest clip and trim it with , or ask for several events in one clip ("three explosions, one after another") and split them with . Extending the other way, past a , is the last-frame chain under Editing existing footage.

Two lanes, told apart by the input in hand. Footage plus a new track is with and the ask in the user's words: these members redraw the mouth region and keep the rest of the frame ( and , or and ; some take typed with a voice instead of audio, never both). A still plus a track is : the talking-avatar members animate the whole portrait and invent its motion, so they serve a portrait with no footage, or footage whose motion is disposable, and never rescue a clip whose lipsync failed, since the result is a different performance of a different shot. Naming lipsync as the capability gets re-read as one of these two, so name the input.

Preflight is free and decides the run. Sync members redraw pixels around the mouth they detect; a face that is small in frame, turned away, occluded, or blurred by motion gives them nothing to redraw, and the run then completes and bills with the mouth still under the new track, the failure users report as "no lipsync at all". the clip's (free, from ): the lips should read at display size. When they do not, crop onto the face first (, the compositor teaches and a fixed first-party id like the cut tools above, or a generative reframe prompted to tighten on the speaker; an outpainting reframe widens the frame and shrinks the face) rather than retrying members. A crop lowers resolution, so read the cropped clip's width and height off against the member's floor before pricing the sync run, and upscale it or pick a member without a floor when it falls short. With several faces in frame, pick a member exposing speaker selection ( on one at authoring time) or crop to one speaker. Per-member limits sit in the schema, not the description: one member takes clips of 2 to 10 seconds at 720p to 1080p, others carry none; the field carries , so trim to the shipped range before the run (Duration limits above).

Plan gating is common here: flags such members on their ranked entry with and names the plan in (unless its response says plan gating is ), and running one anyway returns a 403 naming the model and the plan it needs, which no retry clears; pick an available member or surface the upgrade, per the skill's plan row.

A completed job proves the model ran, not that the mouth moved. Before shipping, download the output and the source and compare frames at the same timestamps (sweep both into contact sheets locally, , and read the mouth region): unchanged mouth pixels across the sweep mean the face was not found, and the same payload reproduces it, so change the framing or the lane rather than the seed. Listen as well: whether the input's own track survives is undocumented (Dubbing above).

with and the user's words finds members that turn a finished song into a music video in one call. The whole-song member live at authoring time took 10 seconds to 6 minutes of and returned a video as long as the track (the same ranking also lists segment-length members capped near 20 seconds), so:

Pick this lane for one pass over the whole track; beat-cut shots under per-shot direction are . Launch with and , since the job scales with the song.