movie maker · skill pública Open Skill

movie maker

Movie Maker is a director-grade, script-to-screen AI pipeline for producing professional cinematic short-form video. It transforms a one-line idea, a draft script, or a character photo into a complete, coherent film — with a natural beginning, rising action, and satisfying ending — while keeping character, location, and prop identity locked across every cut. What it does Script stage: Builds a Story Plan and Beat Budget with a clear goal, question, and beat structure, so every scene moves the story forward. Planning: Locks a world setting, a bold visual style brief (medium, palette, lighting, mood), and an entity list (characters, locations, props) with identity goals. Reference assets: Generates or binds reference images for characters, environments, and props — user-uploaded photos become the lead's locked identity. Render: Shoots 3–6 timed shots per clip via reference-to-video (ref2v), with explicit cinematography, Arabic (or any language) dialogue, realistic diegetic sound effects, and a matching non-diegetic score. Post: Merges multiple segments into one final film (landscape 16:9 or portrait 9:16) with a cinematic music score layered underneath. Use cases Cinematic short films and micro-series (episode 1 or continuing episodes) Trailers, promos, and mood pieces Action, drama, thriller, romance, comedy — any genre Character-driven stories where a specific face/photo is the hero Output specs 10-second reference-bound clips (default), combinable into longer films (e.g., 90s = 9 × 10s) 720P resolution, portrait or landscape Self-contained deliverable: final video + music track One-stage edits supported: rewrite the script, restyle a character, change the look, re-render, or continue the story — without rebuilding the whole pipeline.

@tefa4story

Film

72 9 57

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/898887/2026-09-02/pa5k87dt.jpg

# Movie Maker (Reference-to-Video)

**Director-grade script-to-screen agent** for short-form cinematic video: one coherent clip where characters, locations, and props stay locked across cuts. Default **10s ref2v** per clip.

## When to use

- User wants a cinematic short film / beat / micro-series / trailer from an idea, a script, or reference images (including a user photo as the lead).
- User wants to revise one stage: rewrite the script, restyle a character, change shots, or re-render.

## Unique flow (every run)

If source is missing → `search_web` (max 2) → write **`## Story plan`** + **`## Beat budget`** as visible assistant markdown → plan → refs → timed prompt → video. This is **the gate**: no `generate_image` / `generate_video` before both blocks exist.

## Prompt language

- Tool-facing prompt text is **English** (subject definitions, style brief, shot intentions).
- Dialogue inside the prompt keeps the script's language tag: `<d>[Arabic] ...</d>`.

## Stage 1 — Scriptwriting

Write two visible markdown blocks (not a paraphrase):

```markdown
## Story plan
Source: agent-invented | user outline | user script
Fidelity: faithful | verbatim | inspired
Route: (3–5 phases; direction only)
Window: (one contiguous scene-chain — same want, same space/time)
Question: (one outcome clause answerable in this clip)
Cursor: (the deed that answers Question at camera scale)
Last Button: (drawable prior last frame; "none" before Ep1)
Deferred: (the next source-order question at camera scale)

## Beat budget
Format: episode | trailer | showreel | mood
Story: [person] wants [one visible thing] but [one visible obstacle], so this clip they [Cursor deed]
Scene goal: (must equal Story plan Cursor)
Assigned: (proof increments for Scene goal only; ≤4)
Deferred: (must equal Story plan Deferred)
Button: (visible last frame that proves Question)
```

Rules: 10s per clip; 2–4 short dialogue lines max; Assigned expands Cursor only; never compress/skip with "suddenly/days later".

## Stage 2 — Planning

- `world_setting`: civilization, cultural geography, period, world type, `cultural_prompt_anchor` (English, 15–30 words).
- `visual_style_brief`: **English, 50–120 words, style only** — medium, palette, lighting, mood. Commit to a bold direction.
- Entities: characters / scenes / props with identity goals. User uploads bind directly to entities — **skip regeneration** for those.
- Shot plan: 3–6 shots per clip; every shot names framing + camera movement + cut motivation; dialogue fits 10s.

## Stage 3 — Reference assets

```text
generate_image(
  prompt=<expanded English paragraph>,
  model='t2i',
  orientation='portrait'|'landscape'|'square',
  resolution='720P',
  name=<entity display name>,
  entity_type='characters'|'locations'|'props'
)
```

- Exactly **one** reference image per entity. Never put character names inside prompts.
- User uploads bound to an entity → pass the upload URI straight to ref2v.
- Picture numbering order in ref2v: characters → locations → props.

## Stage 4 — Render (ref2v six-part prompt)

`generate_video` requires **full six-part prompt**, in order:

1. `subject_definitions:` `<Subject N>` is ... in `<Picture N>` (identity only).
2. `summary:` `[reference generation]` + how subjects unfold across shots.
3. `retention_analysis:` per subject — `fully_preserved` (identity kept; reference composition NOT copied).
4. `detailed_description:` style opening + `[Shot 1]..[Shot N]` with explicit camera work; dialogue in `<d>[Language] ...</d>`; cuts usually at 00:03.333 / 00:06.667.
5. `overall_soundscape:` ambient + diegetic SFX (gunfire, footsteps, explosions).
6. `non_diegetic_music:` score heard by audience.

### Call (all fields required)

```text
generate_video(
  prompt=<full six-part text>,
  input_urls=[<Picture N> URIs in order, at least 1],
  model='ref2v',
  gen_mode='reference_to_video',
  duration=10,
  orientation='portrait'|'landscape',
  resolution='720P',
  name=<episode title>,
  entity_type='final_videos'   #