Generate YouTube B-roll concepts with MCP

Turn the edit need into one B-roll shot

Start from the sentence the footage must support, not from a generic request for “cinematic B-roll.” Write a one-line edit purpose such as: “Cover the transition from manual planning to an automated creative workflow.” Then choose one visible action that communicates that idea without spoken dialogue. A desk-level tracking shot past paper notes toward a laptop showing an abstract workflow is narrow enough for a short generated clip; a request for an entire creator journey is not.

Plan this as a `16:9` insert for a landscape YouTube edit. YouTube identifies `16:9` as the standard aspect ratio on computers and says its player adapts to other native ratios, so landscape is a deliberate delivery choice rather than a claim that every YouTube video must use it. OfflineCreator's public MCP `generate` schema accepts `16:9`, `9:16`, or `1:1`, defaults to `16:9`, and does not publish a duration field.

Narration beat

One sentence or edit transitionWrite what the viewer should understand while the insert is on screen.

Visible action

One subject action plus one restrained camera moveAvoid asking a short clip to contain several scenes, cuts, locations, or transformations.

Delivery frame

16:9 landscapeKeep important action away from the extreme edges so the editor has room for reframing and overlays.

Choose a five- or eight-second concept before spending

Call `list_models` and `get_credits` immediately before the paid step. The current public Studio catalog lists Kling 2.6 Pro as a 42-credit text-to-video model for social video and Veo 3.1 Fast as a 96-credit text-to-video model for premium campaigns. Those labels and costs are a dated catalog snapshot, not a quality ranking or a promise about which model will fit a particular channel.

For this guide's bounded example, use a five-second plan on the current Kling path or an eight-second plan on the current Veo path. fal's Kling 2.6 Pro text-to-video documentation exposes `16:9`, `9:16`, and `1:1` with five- or ten-second output. Its Veo 3.1 Fast documentation exposes `16:9` and `9:16` with four-, six-, or eight-second output. These are provider capabilities. OfflineCreator's published MCP schema still has no caller-supplied duration argument, so confirm the Studio model record and do not add an unsupported field to the tool call.

Kling concept

Five seconds, 16:9Use one compact action and preserve a brief stable frame for the edit point.

Veo concept

Eight seconds, 16:9Use the extra time for a slower reveal or longer handle, not for several unrelated shots.

Preflight rule

Re-read model and credit dataThe live Studio response controls the paid decision; this page does not override it.

Write a shot specification instead of a mood list

Use six prompt parts in a fixed order: subject, setting, action, camera, light, and exit frame. Example: `16:9 close desk scene at blue hour; three paper storyboard cards in the foreground and an unbranded laptop beyond them; a hand slides the final card into place while a simple node diagram comes into focus on screen; one slow lateral camera move; cool window light with a warm desk-lamp edge; final second settles with the cards and laptop sharp; no words, logos, jump cuts, or extra hands.`

This prompt describes one editorial insert, not a finished YouTube sequence. It does not ask the model to invent exact interface text, produce a talking presenter, cut between angles, or time itself to narration. If using the five-second plan, keep the slide and camera move short. If the current model path produces eight seconds, extend the settled handle instead of introducing a second scene. The editor can trim a stable end; an overloaded generation cannot be cleanly simplified after the fact.

Plan a three-shot package without asking for a montage

When one insert is not enough, write three independent generation briefs and review each as its own shot: an establishing view, a process detail, and a closing texture or transition. Keep visual continuity explicit by repeating the approved palette, time of day, environment, and material cues in every prompt. Do not claim character, object, or interface consistency merely because the wording is repeated; generated shots still need a human continuity review.

A useful package for the workflow example is: shot A, a wide desk environment with the cards and laptop; shot B, the close card-slide action from the prompt above; shot C, an abstract close-up of warm light moving across the card edges with a clean fade area. Generate and label them separately as `A01`, `B01`, and `C01`. Record the actual returned duration rather than assuming every file matches the editorial plan.

A — establish

Environment and subject relationshipGive the viewer orientation without requiring readable generated interface text.

B — explain

The one action tied to narrationThis is the essential insert; generate it first if the credit budget permits only one attempt.

C — transition

Texture, light, or a stable closing frameUse it as optional edit coverage, not as evidence that the model created a coherent montage.

Run the MCP sequence with an explicit human gate

Use the published read-only tools before allowing a paid call: `list_models` returns Studio launch models with credit costs and workflows, while `get_credits` returns the current account credit balance. Choose a returned text-to-video model, then show the operator its displayed cost, the `16:9` ratio, the prompt, and a one-generation limit before requesting approval. After approval, call `generate` with `modelId`, `prompt`, `aspectRatio: "16:9"`, and a deliberate `wait` choice. Do not send `duration` or an audio setting because neither appears in the public `generate` input schema.

The public server description says text-to-video jobs queue immediately. In its handler, `wait` omitted or false returns the started generation response; `wait: true` polls a returned generation ID for completion or failure for up to 120 seconds. For a separately polled job, retain the generation ID and use `wait_generation`. Once complete, `download_output` returns a short-lived signed URL without streaming media bytes into model context. Retrieve the file while the link is valid, log the actual result, and stop for review before authorizing a revision.

Choose the next guide from the review result

Use the one-shot planning guide when the purpose, first frame, or motion remains ambiguous. Move to the product-motion guide when an approved still must anchor product appearance rather than starting from text alone. Return to the MCP workflow directory when the deliverable changes from YouTube B-roll to another asset class. These links follow the next production decision instead of repeating the same landscape-shot instructions.

Keep the draft inside the evidence boundary

This page establishes a published `16:9` MCP input path and documents provider duration options, then supplies an editorial method for planning one B-roll insert. Those sources define available controls and model capabilities; they do not establish output quality, prompt reliability, continuity across separately generated shots, audience retention, production speed, or model superiority. Treat the example prompt as a shot specification to review, not as evidence that a model will reproduce its intended composition or motion.

Keep each generation independently reviewable: record the returned model, ratio, observed duration, prompt revision, and generation ID, then inspect the actual clip before it enters an edit. Recheck the live model list and credit cost before spending, and recheck the MCP schema plus provider ratio and duration documentation under the quarterly freshness cadence. Human approval remains the decision gate for another paid attempt and for publication; this workflow does not convert documented availability into a creative-performance promise.