andreakeating1982/video-generation

Video generation

Make the video the user asked for as a playable file or link. Two ways to make the picture, often combined:

• Model generation with media_generate_video: people, products, places, cinematic camera work, stylised or 3D animation, transformations.
• Code rendering in the sandbox: anything that must be exact, such as text, logos, numbers, charts and UI. Frames are drawn in HTML or Python and encoded with ffmpeg.

ffmpeg in the sandbox joins, trims, subtitles, mixes and exports.

Start here

1. Work out what the video must communicate, for whom and where it will play. Settle the aspect ratio (landscape 16:9 for YouTube and desktop, portrait 9:16 for Reels, TikTok and Shorts), the approximate length, and which assets already exist (photos, logos, footage, voice). Use what the request implies and ask only when a missing answer would change the result.
2. Pick the route in the table below and read its files in one parallel batch. Skip files that the task does not need.
3. Plan before generating: one central idea, then a shot or beat list with what happens, the camera, the sound and the length of each part.
4. Generate, inspect, assemble if needed, deliver.

Request	Read
Motion graphics, kinetic type, logo stings, animated data, UI animation, design showreels	workflows/motion-graphics.md, then references/code-rendering.md for code renders or references/video-prompts.md for model renders
Ads, product promos, launch teasers, UGC-style clips, social hooks	workflows/ads-and-promos.md and references/video-prompts.md
Short films, music videos, recurring characters, story clips	workflows/stories-and-characters.md and references/video-prompts.md
The user's own footage: trim, highlights, join, subtitles, music, resize	workflows/editing-footage.md and references/ffmpeg.md
Joining generated clips, adding music or titles	references/ffmpeg.md

The video tool

Reading this skill unlocks the media tools. If media_generate_video is still missing, call media_generate with a one-line brief. Video is not available on the Free plan.

• The tool's description lists every model with its lengths, resolutions and credits per second, and says which model fits which request. Use the model the user names; otherwise follow that guidance.
• MiniMax H3 Max (the default) returns the finished videoUrl in the same call. WAN 3.0, Kling 4.0 and the Seedance models render in the background: the call returns a jobId at once and you are resumed automatically with the video after a few minutes. Never poll and never call the tool again for the same clip. When a plan needs several background clips, start them all in the same step so you are resumed once.
• H3 Max and Seedance 2.5 can follow a multi-shot prompt (Shot 1, Shot 2, ...) inside one clip. WAN 3.0 renders one continuous take with strong camera work.
• Every model generates audio with the picture. Describe dialogue (quoted), ambience, sound effects and music in the prompt.
• Modes: text-to-video; image-to-video (keyframes.first); first-to-last frame (keyframes.first + keyframes.last); reference images (cited as "Image 1", "Image 2" in list order — @Image1, @Image2 with Kling 4.0 — not combinable with keyframes.last). An image start frame gives better, more controllable results than text alone.

Rules that decide quality

• People and products need image references. Use the user's photos or generate a reference with media_generate_image first, then pass it as keyframes.first or in references. Settings, lighting and style can be described in words.
• Exact text, logos, prices and UI are never left to the video model. It garbles them. Render them in code or overlay them with ffmpeg.
• One clip per request whenever it fits the model. Split only for real limits (a 45-second ad is several clips joined with ffmpeg) or when the user wants separate clips.
• Write the model prompt in English. Quote spoken lines and on-screen words in the language the user wants.
• Spend deliberately. For several clips or long background renders, state the shot plan and the total credits in one line. Ask for a go-ahead only when the total is far above what the user would expect.

Check and deliver

• Inspect every result before delivering. Use ffprobe for duration, size and audio, and a contact sheet of frames (see references/ffmpeg.md) viewed with file_view. file_view also analyzes a video file or URL directly when you need to check motion, speech or sound.
• Never generate a clip again on your own. Every clip costs credits. If a result has a real failure (wrong subject, broken identity, garbled required text, a missing action), show it, say in one sentence what is wrong, and ask whether to make a corrected version. Generate again without asking only when the user told you to keep fixing it until it is right.
• Code renders and ffmpeg edits cost no credits: fix real problems in them yourself, but stop polishing once a frame works.
• Delivery: a single model clip is shown in the chat from its videoUrl. Files you assembled or rendered in the sandbox go out with send_attachment. In one or two lines, state the format, aspect ratio, length and whether there is sound.
• For a narrated explainer, a talking avatar or long timeline editing, the platform's Video Agent, Talking Videos or Video Editor can fit better. Offer the link when it does.