Mission
Use a supplied movement reference or constrained visual endpoints to make motion intentional and inspectable. This skill covers first/last-frame transitions, a last-image continuation control, and model-specific driving video or motion-control workflows. It does not promise exact physical tracking, camera telemetry, or a universal camera-path interface.
Required inputs
- subject or product reference and rights/consent state;
- driving video, first frame, last frame, or motion intent;
- action, timing, camera behavior, screen direction, and preservation rules;
- target duration, ratio, resolution, audio behavior, and delivery format; and
- tolerance for creative interpretation, variants, cost, and review owner.
Clarify whether the reference supplies appearance, motion, camera movement, or all three. Do not treat a driving clip as permission to copy an unrelated person's identity or performance.
Select the motion path
| Intent | Path | Schema discipline |
|---|---|---|
| Controlled transition between visual endpoints | first/last-frame model path | preserve list order exactly: first, then last |
| Continue an existing shot toward a specified image | I2V with last_image where supported |
verify the selected endpoint accepts the field |
| Transfer movement from a driving video | Kling motion-control or another live endpoint | required image/video roles and audio defaults vary |
| Ordinary camera/action direction | T2V or I2V prompt | describe observable movement; do not invent unsupported controls |
Use MODELS.md and the live catalog. Current reference shapes include Seedance first/last-frame and omni-reference endpoints, Veo I2V with an optional last image, and Kling motion control with image/video inputs. The exact prompt syntax, duration, ratios, audio behavior, and reference limits are endpoint-specific.
Execute
- Assign stable roles and upload local references. Record URL, order, source, rights, and whether original audio should remain.
- Select one motion path and write a contract for subject, movement, camera, timing, preserve/avoid rules, and acceptance checks.
- Submit a small proof first when the motion is costly or uncertain. Retain
prediction ID and poll with
media.check_result. - Review trajectory, limbs, contact points, camera direction, identity, product geometry, temporal stability, and audio. Compare with the driving reference only for the intended motion, not for unauthorized likeness.
- Lock the approved parent before dependent edits, crops, captions, or upscaling. Report any model field that was unavailable rather than approximating it silently.
Do not hardcode a camera_path, extract_frame, or arbitrary motion field
unless the selected live schema exposes it. For frame extraction, use a
separate approved media utility; video generation alone is not an extraction
contract.
Output contract
Return motion path, reference-role table, model/endpoint, exact schema fields,
prediction ID, output URL, status, audio treatment, QA results, and failed or
unsupported controls. Use video QA and save
permitted receipts under .video/ via
reports-and-receipts.md.
Generation is draft-only. No physical-world claim or publication action is implied by a generated motion clip.