Models / Runway / Video to video

Runway video to video prompting guide

How to prompt Runway in video-to-video mode. Every point links its source; community tips are marked. Your agent gets the same guide through the Atlas MCP (get_guide).

Prompt structure

edit instruction→what to keep→keyframe plan

• T2V: '[Camera] shot of [subject] [action] in [environment]. [Supporting descriptions]'. Structure matters less than removing ambiguity. source ↗

What the text should describe

• Aleph 2.0 (aleph2) on /v1/video_to_video: videoUri + promptText (optional, ≤1000 chars) + 1–5 keyframes, each an image placed by `at` (fraction 0–1 of the source) or `seconds`, with an optional range {start_seconds,end_seconds} to confine the edit window. source ↗
• Act-Two (act_two, /v1/character_performance) transfers a 3–30 s driving performance video onto a character image or video; the character needs a clearly visible face that stays in frame. source ↗
• Creators use Aleph 2.0 keyframes to pin specific moments and let the edit carry through the scene, instead of relying on one prompt. community source ↗

What not to describe

• Don't mix keyframes with and without range: all keyframes must set a range or none may. source ↗

Camera vocabulary

• For still shots: 'The locked-off camera remains perfectly still.' source ↗

Responds well to

• Sequential prompting: 'X, then Y. Finally, Z.' or '[00:01] X. [00:03] Y.' source ↗

Duration and limits

• Chain longer sequences: last frame of one generation as the next image input. source ↗

Avoid and failure modes

• Positive, concrete phrasing; no conflicting instructions; very long multi-paragraph prompts over-constrain. source ↗
• Unwanted cuts = duration too short; add 'Continuous, seamless shot'. source ↗

Negative prompt

Supported via .

Next

Sources

• academy.runwayml.com/guides/prompting-guide
• docs.dev.runwayml.com/api/
• x.com/kimcarldaniel/status/2068332851162955888