Veo, Google's video family on Scenario, ships four members at authoring time: Veo 3.1 (full quality, reference images with a style switch), Veo 3.1 Fast (same modes, quicker, no switch), Veo 3.1 Lite (frame anchors only, cheapest), and Veo 3.1 Extend Video (continues a clip). Discover them with search and treat model_schema_get as the contract: members agree on parameter names and disagree on which inputs exist at all.
Connection and the core loop: see the scenario skill in this repo; model-agnostic video work: the scenario-video skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.
Mode follows from the inputs (names from the live schema):
\n| Mode | \nInputs | \nBehavior | \n
|---|---|---|
| Text | \nprompt | \naspectRatio 16:9 or 9:16 | \n
| First frame | \nimage (+ prompt) | \nopens on that frame; a mismatched image may be cropped | \n
| Transition | \nimage + lastFrameImage | \nbridges the stills; prompt one continuous move | \n
| Reference | \nreferenceImages (1 to 3) | \nsubject or style consistency; duration locked to 8 | \n
| Extend | \nvideo (Extend member) | \ncontinues the clip; 16:9 source, short side 720p or 1080p | \n
image and referenceImages are mutually exclusive. referenceImagesType exists on full Veo 3.1 alone and reads the whole array one way: ASSET (default) carries subjects, objects, and scenes; STYLE carries palette, lighting, and texture. Fast takes references without the switch; Lite takes none. Shared across the three generators at authoring time: negativePrompt, resolution (720p, 1080p), duration (4, 6, or 8 seconds), and seed. generateAudio is required on every member and moves the price. The Extend member takes only prompt, video, generateAudio, and seed: no duration, resolution, or ratio controls. At authoring time price spanned roughly five times between Lite and full 3.1 for one image-to-video job, and one extension cost more than a fresh full generation, so aim for the shot in one 8 second pass and dry_run the same payload on two members before a batch.
With generateAudio: true, Veo renders synchronized dialogue, effects, and ambience, and it takes audio direction literally. Put spoken lines in quotation marks, prefix effects with SFX: and room tone with Ambient noise:. A prompt with no audio direction still gets a soundtrack, just not the one you meant. When a clip needs silence for later scoring, set generateAudio: false rather than prompting for quiet.
Write present-tense prose ordered as cinematography, subject, action, context, style: camera and shot scale first, one primary arc, roughly 90 to 250 words. For several beats in one clip, timestamp lines work ([00:00-00:02] Medium shot...), spans fitting inside the chosen duration.
search with target=\"models\", query=\"veo\", public=true. Prefer the newest non-deprecated hits, e.g. model_veo3-1 and model_veo3-1-fast (live hits at authoring time: re-discover each session).model_schema_get on the chosen id: confirm which reference inputs exist before writing the payload.upload_asset two character stills (see the scenario skill) to get asset ids.model_run with that model_id, dry_run=true, and parameters={\"prompt\": \"Medium shot, the knight lowers her visor and says, \\\"Hold the line.\\\" Torchlight flickers on wet stone. Ambient noise: distant thunder. SFX: metal visor clank.\", \"referenceImages\": [\"asset_a\", \"asset_b\"], \"referenceImagesType\": \"ASSET\", \"duration\": 8, \"resolution\": \"1080p\", \"aspectRatio\": \"16:9\", \"generateAudio\": true}. Estimate the same job on the Fast id (drop referenceImagesType, absent there) and compare.model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.asset_display the output and review with sound. To iterate on wording alone, hold an explicit seed fixed across reruns.image with referenceImages: mutually exclusive; pick the opening state or consistency.referenceImages: reference mode supports only 8 seconds.referenceImages to Lite or referenceImagesType to Fast: the schema decides which inputs exist.generateAudio: it is required on every member; pass it explicitly.negativePrompt for what to discourage.