Movie Maker (New) · Open Skill 公開スキル

Movie Maker (New)

Director-Grade · Script to Screen · ref2v Native The agent skill for creators who want a finished 10-second cinematic clip—not a pile of disconnected frames Movie Maker is a production-native Agent Skill for Digen. It orchestrates a complete reference-to-video (ref2v) pipeline: Script → world & shot plan → reference assets → six-part ref2v prompt → 10s render No storyboard grids. No shot-by-shot image chains. One coherent short-form clip where characters, locations, and props stay locked across motion and cuts.

@sheikhsd

Film

22 0 6

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4125068/2026-10-19/matfp2fh.jpg

---

allowed-tools:
  - search_web
  - list_models
  - generate_image
  - generate_video
  - mix_clips
  - recognize_image
  - create_document
  - get_document
  - update_document
  - memory_query
  - memory_store
  - memory_whoami
---

# Director-Grade · Script to Screen · ref2v Native

A production-native agent skill for creators who want **one finished 10-second cinematic clip** — not a pile of disconnected frames.

Pipeline: **Script → world & shot plan → reference assets → six-part ref2v prompt → 10s render.**

No storyboard grids. No shot-by-shot image chains. Characters, locations, and props stay locked across motion and cuts.

**Docs language:** this file and all `references/` are English-only. Match the user's language in replies; recognize intent in any language.

## When to use

- A single cinematic 10s clip from an idea, a script, a logline, or reference photos.
- A short series (Ep 1, Ep 2, …) produced as separate clips in one thread, with identity continuity.
- Reference-driven work: user uploads of characters / locations / props that must stay recognizable.
- Surgical re-renders: change a shot, the camera, the style, the orientation, the dialogue — without rerunning the whole stack.
- A promo, teaser, slice-of-life beat, or character moment for TikTok / Reels / Shorts.

**Not the fit:** storyboard grids, NLE-style editing of long footage, a single multi-minute render, or a video with no creative brief.

## Defaults (user can override any)

| Item | Default |
|---|---|
| Clip length | **10 s** (`duration=10`) |
| Videos per episode | **1** ref2v clip |
| Shots per clip | **3–6** (3–4 simple, 5–6 dense) |
| Orientation | `portrait` (9:16); `landscape` on request |
| Resolution | `resolution='720P'` on every image and video call |
| Video channel | `generate_video(model='ref2v')` — **only** |
| Storyboard images | **Never generated** — shots live inside the ref2v prompt |
| Reference images per render | up to 9 (`input_urls` max 9) |

## Stages

First production read is **only** `references/script-and-plan.md`. Do not load later files in parallel.

| Stage | Read | Tools | Output |
|---|---|---|---|
| 0. Source | — | `search_web` | Source notes (≤5 lines) when no script/brief exists |
| 1. Script & plan | `references/script-and-plan.md` | `search_web`, `create_document` | Beat, one-line Window, 3–6 shot intentions |
| 2. Reference assets | `references/asset-design.md` | `generate_image`, `recognize_image` | character / location / prop refs |
| 3. ref2v prompt | `references/ref2v-prompt-contract.md` + `references/shot-design.md` | — | the six-part timed prompt |
| 4. Final render | — | `generate_video`, `mix_clips` | the 10s clip on canvas |

Use `write_todos` to track stages; at most one `in_progress` at a time. Update the whole table as you go.

### Stage 0 — Source (only when the brief is thin)

Skip when the thread already has a script, outline, or clip-ready text, or the user says not to search. Otherwise run at most **2** `search_web` calls (`max_results` ≤ 3). Never paste raw snippets — write **Source notes (≤5 lines)** with one local chain: `current scene → direct consequence → next adjacent scene`.

### Stage 1 — Script & plan (the gate)

Before any `generate_image` / `generate_video`, the thread must contain:

- `## Beat` — who wants what, and what changes (or the tease / montage line / mood shift).
- `## Shot plan` — 3–6 numbered shot intentions, each with a camera idea and the time window it occupies.

A one-line summary does **not** satisfy the gate. These two blocks are the deliverable of this stage — write them as assistant text. `references/script-and-plan.md` holds the craft (beat structure, shot intention format, series Window/Cursor/Deferred).

Plan the world too: location, time of day, weather, palette, lens set, and the two or three entities that must be visually locked.

### Stage 2 — Reference assets

For each entity that must stay consistent, produce **one** reference image and bind its URI f