Video Generation Studio

/video-generation

Read the request before reaching for anything. Three different things get called "make me a video", and two of them are live capabilities with very different costs. Route first.

1. A NEW clip, invented from a description → generate it

The product generates short clips from text, optionally using a still image as the first frame. This is real and it is paid per second.

• The available lengths, the tiers, and the exact per-second price are stated in your system prompt. Read them there. Do not quote a price, a length or an engine name from this file or from memory — the engines and prices move, and a stale quote is a lie about money.
• Clips are short — a few seconds. This is the single most important fact for routing: nothing a user would recognise as "a video" (an ad, an explainer, a reel, an intro) fits in one generation.
• Over the spend threshold the product asks the user to confirm before it bills. Let that happen; do not talk the user out of it or around it.

2. Anything longer, or anything built from footage that exists → edit_video

Call edit_video. It assembles a timeline from clips, stills, music and narration already in the conversation — generated or uploaded — and it is what turns material into a finished piece:

• join clips in order, trim each one to the part that matters, split, reorder
• transitions between shots; on-screen titles and captions burned in
• retime (speed up / slow down), set per-clip volume, lay a music or voice bed under the whole thing with fades
• reframe: pick a vertical canvas for a reel or a short, letterbox or fill

It writes a document and opens an editor the user can also drag by hand. It does not render the mp4 — the user presses Export for that. Editing itself is free; the export is compute and is charged once, when they ask for it.

The common shape of a real request is BOTH, in order: generate the two or three shots you need, then call edit_video once to cut them together. "Make me a 30-second promo" is not one generation — it is several short ones plus an assembly. Say what you are doing, generate, then assemble.

Footage is named, never keyed: refer to it the way the user does ("the drone shot", "intro.mp4"). Send every step of one edit as a single ops array — if any op is wrong, nothing is saved and you are told which.

3. What genuinely does not exist: making a clip LOOK better

There is no video-to-video picture model. Three operations are not available, and no amount of assembly substitutes for them:

• upscaling / raising resolution
• enhancing / sharpening / restoring detail
• denoising / deblurring / removing compression artefacts

Say so plainly, in one sentence, and never generate a new clip and present it as an improved version of the user's own footage — that is the specific failure this boundary exists to prevent, and it bills the user for something they did not ask for. What you can offer instead: watch the clip and describe it, cut it, caption it, reframe it, or generate fresh footage at a higher tier if they want new material rather than a repair of the old.

Optional: writing a reusable shot spec

When the user wants prompt-writing help — for our generator or an external one — produce a tight spec. Only include relevant lines; make implicit details explicit without inventing new creative requirements.

Use case: <where the clip will be used>
Primary request: <user's main prompt>
Scene/background: <location, time of day, atmosphere>
Subject: <main subject>
Action: <single clear action>
Camera: <shot type, angle, motion>
Lighting/mood: <lighting + mood>
Color palette: <3-5 color anchors>
Style/format: <film/animation/format cues>
Timing/beats: <counts or beats>
Text (verbatim): "<exact text>"
Avoid: <negative constraints>

Prompting best practices: one main action + one camera move per shot; use counts/beats for timing; keep on-screen text short with a locked-off camera; shorter clips follow instructions better; add a brief "avoid" line when artifacts (flicker, jitter, warped hands) appear. On-screen text is usually more reliable burned in with edit_video than asked of the generator.

Guardrails

• Never fake a result: no invented clip, no JSON "video" object, no placeholder link. If something cannot be done, name it and move on — no apology spiral.
• Never quote a price, duration or engine name from this file; the system prompt is the only current source.
• No real people or public figures, no copyrighted characters or music, general-audience content only.

Usage

/video-generation $ARGUMENTS