---
name: video-framework
description: Unified vclaw-native front door for creating videos by routing across copy, create, narrated, presentation, long-form, film, and UGC workflows over the canonical `vclaw video` commands.
---

<Purpose>
Video Framework is the flagship vclaw-native video production surface. It gives users one place to start — “make me a video” — and then internally routes to the appropriate workflow mode while keeping the public experience inside the `vclaw` CLI.
</Purpose>

<Use_When>
- The user wants to make a video, copy an ad, generate a product video, create a narrated explainer, restyle a presentation, or run a longer-form cinematic workflow
- The user wants a unified vclaw-native experience instead of raw Python/Bun scripts
- The task should run through the canonical `vclaw video` pipeline commands
</Use_When>

<Do_Not_Use_When>
- The user only wants a single static image
- The user wants low-level backend debugging on a specific legacy script without the vclaw product surface
- The user is asking for a generic non-video orchestration task
</Do_Not_Use_When>

<Current_Product_Boundary>
This skill is the new front door. It should:
- classify requests into internal modes such as COPY, CREATE, COPY NARRATED, PRESENTATION, LONG-FORM, FILM, and UGC
- gather missing inputs and preferences
- prefer vclaw-native orchestration and the on-disk `projects/<slug>/` state
- drive the canonical `vclaw video` commands instead of exposing raw scripts directly

Initial migration direction:
- `veo-cli` is treated as a backend/service adapter source
- the imported `skills/video-replicator/` tree is treated as the legacy
  workflow/phase-engine reference set
- the public UX should not leak `.claude` path assumptions, printed MCP commands, or raw Python/Bun entrypoints
- CLI planning surface: prefer `vclaw studio --dry-run` when the user wants a guided menu-like experience. Studio should produce the command plan; this skill remains the agent reasoning layer that interprets creative intent.

Reference guides:
- `references/checkpoint-protocol.md`
- `references/stage-directors.md`
- repo-local follow-on skills:
  - `skills/video-analyze-template/SKILL.md`
  - `skills/video-clone-ad/SKILL.md`
  - `skills/video-storyboard/SKILL.md`
  - `skills/character-creator/SKILL.md` — REQUIRED SUB-SKILL for any cast/character work (drives `vclaw video character-auto-create`)
</Current_Product_Boundary>

<Example_Requests>
- "make me a product ad video from these assets"
- "copy this ad with a new subject"
- "create a narrated explainer video"
- "turn this presentation into an animated video"
- "start a cinematic film workflow for this concept"
</Example_Requests>

## Intake and execution

Read the existing project and the user's brief first. Ask only for missing
information that materially changes the output: subject/brand, audience,
look, supplied assets, duration, delivery format and audio. Preserve explicit
choices and previous authorisation; do not force a seven-question interview.

Use `vclaw schema --json` and `vclaw studio --dry-run` to discover current
commands. Generic execution quality and the Flow model are separate settings:

```bash
vclaw video brief --project <slug> --title "<title>" --intent "<intent>" \
  --aspect-ratio 16:9 --quality fast --audio off
# Optional Flow-specific choice, when that route is intended:
vclaw video set-execution-profile --project <slug> --veo-model omni-flash
```

`--quality` accepts `fast|quality`. Flow choices such as `lite`, `free` and
`omni-flash` belong to `--veo-model`, not `--quality`. Provider routing is
resolved from project state and current readiness; do not translate old
backend aliases into invented CLI flags.

Show the proposed storyboard and actual render contract before the first paid
image, audio or video call. Confirm scope and estimated spend once, then
continue within that authorisation. A material change in cost, route or scope
needs renewed consent. A dry-run or enqueue result is not a rendered preview.

## Dispatch by deliverable

- Reference breakdown → `video-analyze-template`, then `video-clone-ad`.
- Authored project scenes → `video-storyboard`; cast → `character-creator`.
- Cinematic prompt packets → `ai-filmmaking`; visual blueprint → `ai-director`;
  direct-paste scene direction → `ai-film-director`; narrative production →
  `movie-director`.
- Host/deck video → `brand-presenter` with the requested profile.
- Belief campaign → `ugc`, with a separate project per finished video.
- Motion graphics → `mograph`; existing talking-head overlays → `motion-reel`.
- Review/handoff → `video-production-handoff`; final media packaging →
  `video-post`; portfolio operations → `video-portfolio-ops`.

For direct rendering, plain `execute`/`produce` can submit provider work.
`render-scenes --method <route> --confirm-spend` runs sequentially on that
route; omit `--fallback-chain` when a fixed route is required. Inspect each
returned report and scene result, not just exit status.

`pool`, `produce --auto-chain` and `batch-submit` compile durable queue tasks;
they do not render. Follow the returned task IDs through `cinema-work` and its
applicable authorisation/reconciliation contract. See the current Runway skill
for free exploration and `docs/CLI_REFERENCE.md` for exact paid quotes.

## Audio and completion

State intended dialogue, narration, ambience and music per scene. Do not add a
universal no-music/no-voice tail when the user explicitly wants speech or music.
Keep identity/reference requirements and the chosen route's audio contract.

`narrate`/`dialogue` generate speech; `assemble` builds the final from the chosen
inputs. `remix-narrated` only concatenates existing narrated MP4 clips and is
not a substitute for narration generation or general assembly.

Inspect generated media before using it downstream. Report failed, missing,
substituted and still-pending work explicitly. Publish readiness comes from
current canonical review evidence, not an enqueue receipt or local checkbox.

The older intake references in `references/` preserve creative vocabulary;
verify any embedded command/flag recipe against current schema before use.
