口播视频 B-Roll 画面生成

# Talking-Head B-Roll Generator

Generate shot-specific B-roll stills from a talking-head script or transcript.

## Required Aident Actions

- <action-tag>direct:fal:fal_text_to_image</action-tag> (required inputs: prompt): Run an active Fal text-to-image model. Defaults to fal-ai/nano-banana-2. Call fal_list_text_to_image_models before selecting another model to inspect current pricing, required fields, defaults, and model differences....

Use Aident Loadout to read the current Action schema before constructing inputs. Check the required integration connection in Aident Vault. If an Action is billable or mutating, run Aident preflight, show the affected target and quoted cost or risk, and wait for explicit user confirmation before execution.

## Workflow

1. Confirm the requested outcome, source material, destination, audience, constraints, and acceptance criteria.
2. Apply the source-derived guidance below to create a concrete plan. Resolve ambiguity before invoking an Action.
3. Select only the Aident Actions whose documented effect directly advances the requested outcome. Do not invoke every listed Action by default.
4. Inspect the current schema and prepare the minimum valid input for each selected Action.
5. Preflight each selected Action. Execute it only after any required confirmation, in dependency order, and carry returned IDs or asset URLs into later steps.
6. Verify the returned IDs, URLs, statuses, or artifacts against the acceptance criteria. Report partial completion precisely and do not repeat paid or mutating calls blindly.

## Source-Derived Guidance

- Collect the script or transcript, target platform, aspect ratio, visual style, brand constraints, and the lines that need visual coverage.
- Create a concise shot list that maps each selected line to one concrete visual concept and avoids repetitive literal illustrations.
- Use direct:fal:fal_text_to_image for approved shots, carrying consistent style, palette, subject, and aspect-ratio constraints into each prompt.
- Verify that every returned still supports its assigned line, contains no unintended text or brand misuse, and is visually consistent. Return still asset URLs and shot mapping only; do not claim to have edited a finished video.

## Execution Boundaries

- Treat upstream provider-specific commands as background knowledge only. Execute the workflow through the exact Aident Actions above.
- Never request raw credentials in chat. Use Aident Vault connection flows for required accounts.
- Preserve user-provided wording, brand constraints, rights restrictions, and target identifiers. Do not invent authorization.
- Pass assets by Aident asset ID or supported URL fields. Do not expose caller-local filesystem paths to hosted Actions.
- Stop when a required integration is disconnected, preflight rejects the input, the target is ambiguous, or the user declines a required confirmation.

## Output

Return the execution plan, selected Action names, confirmed targets, Aident result identifiers or asset URLs, verification evidence, and any remaining blocked step.

## Attribution

This Skill adapts the reviewed upstream workflow. See [UPSTREAM.md](UPSTREAM.md) for the pinned source and [LICENSE.txt](LICENSE.txt) for the preserved license.