Mission
Supply visual coverage that supports an approved narration, interview, or edit without overwriting the story or pretending that generated imagery is factual footage. Work from a script, transcript, edit decision list, or explicit gap list. Return clips that an editor can place by scene and time range.
Required inputs
- script/transcript or source edit with timecodes;
- desired visual gaps, subject/product/location, and reference roles;
- target duration, ratio, resolution, style, and audio policy;
- whether coverage should be illustrative, archival, product-accurate, or abstract; and
- rights/consent, claims, budget, number of candidates, and review owner.
If the source is a current event, historical claim, medical topic, public figure, or real-world demonstration, require verified source footage or label every generated shot as illustrative. Do not generate a plausible scene and call it evidence.
Build a coverage matrix
For each gap, record:
coverage_id / source timecode:
narration or edit purpose:
visual subject and action:
literal versus illustrative status:
reference roles and preservation rules:
shot size / camera / light:
duration / ratio / audio:
acceptance check:
Prefer varied coverage—establishing, detail, hands/product interaction, environment, abstract texture, and transition—without changing the intended meaning. Keep one visual idea per clip and avoid random montage filler.
Execute
- Freeze the script/edit parent and rank gaps by narrative importance.
- Upload authorized references or hand off still preparation to the image agent. Record asset roles and provenance.
- Choose T2V or I2V from MODELS.md, inspect the live schema, and generate a small candidate set per gap.
- Poll jobs, review temporal stability and factual implication, and mark each
clip as
literal-source,illustrative-generated, orabstract-generated. - Lock selected B-roll parents. Combine only when the user asks for a rough assembly; otherwise return editor-ready clips with placement notes.
- Use video QA for continuity, product/ likeness, captions, audio, rights, and export checks.
Do not generate a replacement for missing source evidence, hide a failed gap, or silently insert generated footage into an approved edit.
Output contract
Return a gap/coverage table, parent/source IDs, model/endpoint, prediction IDs,
output URLs, literal/illustrative labels, placement notes, QA findings, and
failed candidates. Save permitted manifests and receipts under .video/ using
reports-and-receipts.md.
This skill creates draft coverage, not a claim of what happened in the world.