Veo 3.1音频提示契约

Applicable scenario

Use before sending a Veo 3.1 native-audio generation request that contains dialogue, sound effects, or ambient sound. Python 3.9+ is required; no model key, network call, upload, or external package is required.

Inputs and outputs

Input is a redacted JSON request with model, duration_seconds, aspect_ratio, and ordered audio_cues. Each cue has an id, kind, start and end seconds, and text. Output is veo31_audio_cue_contract.json plus a Markdown review report.

Installation and execution

1. 1. Copy the skill directory or run it in place.
2. Edit a redacted request and keep speaker, brand, and private-person data out of the file.
3. Run python scripts/build_audio_cue_contract.py <request.json> --out-dir <output-dir>.
4. Only a ready_for_generation report may be handed to the model operator; the script never performs that handoff itself.

Rules

The contract accepts Veo 3.1 durations of 4, 6, or 8 seconds, requires a non-empty prompt cue text, rejects duplicate ids, rejects overlaps for cues assigned to the same channel, and requires every cue to fit inside the clip. The cue kinds are dialogue, sfx, and ambient. These are packaging and acceptance rules, not claims that the model will obey every cue.

Recovery

Fix all listed blockers, rerun the script, and preserve the blocked report for review. Do not silently truncate cues or invent missing timing. A provider rejection, quota failure, moderation response, or missing audio track must be handled by the provider\'s normal controls and then reviewed by a person.

Permission and privacy

Do not place API keys, cookies, private voice samples, personally identifying dialogue, or unlicensed media references in the request. The script does not contact Veo, does not upload data, and does not publish results.

Acceptance

The normal example exits zero, emits the expected cue ids and a stable contract hash, and the invalid example exits nonzero with duration and overlap errors. A human must confirm prompt wording, rights, provider settings, and the resulting audio before release.