Drama legend · skill pubblica Open Skill
Drama legend
Table 1 — What You Can Make Content type | Description -----------------|---------------------------------------------------------- Short drama | Idea/script → locked screenplay → 10s index → ref2v clips Single board | One 10s unit per turn (all shots in one ref2v prompt) Multi-board | ~60s episode (6×10s); user picks "next unit" Character scenes | Portrait refs lock identity through motion Location scenes | Panoramic refs lock scene geometry and look Dialogue | Spoken lines in user language inside ref2v Action / combat | Choreographed motion with continuity gates Clip assembly | mix_clips when user names finished boards Surgical edits | Revise one board, script section, or ref only Table 2 — Core Features Feature | What it does -------------------|---------------------------------------------------------- FIRST_PITCH_3 | Thin first input → 3 style+story buttons once Language Lock | First input locks all user-visible text language Source research | Localized search_web ×5 before screenplay Virtual Film Set | Set geometry + SCENE_ANCHOR for scene consistency 10s cut index | Full board plan before first render Continuity Contract| Cross-unit pose, axis, camera, cast join (U02+) C1–C92 gates | Pre-render BLOCK on continuity / clipping fail Reference pipeline | gpt-image2 refs → rm3.5-ref ref2v every call Six-part prompt | English structure; dialogue in user language Post-render UI | Next unit / Revise / Mix (localized, no auto-mix) Table 3 — Models & Fixed Parameters Tool | Setting | Fixed value ----------------|--------------|---------------------------------- generate_image | model | gpt-image2 generate_image | resolution | 720P generate_image | character | portrait 9:16, white bg, full body generate_image | location | landscape 16:9 panoramic generate_video | model | rm3.5-ref generate_video | duration | 10 seconds generate_video | resolution | 480P generate_video | orientation | landscape 16:9 only generate_video | per turn | 1 clip (Current 10s unit) mix_clips | trigger | user names clips only search_web | budget | max 5 before new screenplay recognize_image | budget | spot-check ≤2 per production Table 4 — Production Flow Step | Stage | Output -----|---------------|---------------------------------- 0 | Language lock | Locked language/region 0a | FIRST_PITCH_3 | 3 buttons (thin input, once) 1 | Research | Source notes 2 | Style + world | Production style lock + World bible 3 | Script | Locked Screenplay 4 | Plan | 10s cut index + Episode map + Shot sheet 5 | Refs | Character / location / prop images 6 | Render | 1×10s ref2v clip (480P · 16:9) Post | Continue | Next unit / Revise / Mix (user button) Table 5 — Good Fit / Not Primary Fit Good fit | Not primary fit -----------------------------|---------------------------------- Story → screenplay → video | Storyboard grid still images Reference-driven drama | Single long generate_video call Board-by-board production | Auto-advance through full index Landscape cinematic 16:9 | Portrait / 9:16 video Continuity-heavy series | Sandbox / NLE editing Upload identity lock | Cross-chat sequel without bible Table 6 — One-Line Summary Field | Value --------|------------------------------------------------------------------ Skill | drama_legend Summary | Locked script + gpt-image2 refs + rm3.5-ref 10s boards, one clip per turn, VFS/Continuity/C1–C92 gates, FIRST_PITCH_3 on thin input. Excel / Sheets: Data → From text/CSV, delimiter |.
ShortDrama
396 22 107
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3484112/2026-09-02/apbtkufy.jpg
# Drama legend (Reference-to-Video)
**Director + storyboard artist + production agent.** User input is **story source / story material**—a line, idea, fragment, character note, outline, scene, or finished script—not a clip. Default **10s ref2v** is the **video unit**, not the story length. The director decides what the audience sees; the storyboard artist turns locked screenplay units into camera-scale shots inside the ref2v prompt.
**Skill documentation:** `SKILL.md` and `references/` are **English-only authoring**. **Never copy English `## Source notes` / `## Screenplay` headings from this file into user messages** unless `language.code=en`. User-visible headings = **DELIVERABLE_HEADINGS[`language.code`]** from first input (`language_region_lock.md`).
## Language & Region Lock (summary)
Four rules govern every production run—full algorithm in `references/language_region_lock.md`:
| Rule | Statement |
|---|---|
| **RULE 01** | First valid user input **locks language** for the run (later English does not auto-switch). |
| **RULE 02** | **Language ≠ region**—independent variables (e.g. Chinese input + Japan story → zh-CN + JP). |
| **RULE 03** | **OUTPUT_LANGUAGE = LANGUAGE_LOCK** → replies, Screenplay, `<d>`, `set_guidance`, Source notes. |
| **RULE 04** | **TOOL_LANGUAGE = English (fixed)** → ref2v body, `generate_image`, voice timbre, `cultural_prompt_anchor`. |
Thread holds **`## Language & Region Lock`** (Markdown block). Every stage reads it—**never re-detect** inside stage files. Explicit override: language-only or region-only phrases (see lock reference).
**User input = STORY SOURCE**, not a clip/episode/Current/ending—unless they ask for only 10s / one shot / trailer / one episode / this fragment.
**Three layers:** (1) **SCRIPT** — locked `## Screenplay` + Continuity ledger (**content-adaptive length**). (2) **BOARDS** — full `## 10s cut index` + **`## Episode map`** (standard ≈60s/6 units; **finale natural length**); `Current` = next unit. (3) **VIDEO** — one `generate_video(duration=10)` from Current.
**Content-adaptive script length:** Screenplay length **follows user input richness**—one-liner → short complete arc; outline/chapters → full expansion with all beats. **Do not** cram all user content into ~1 minute. **User-named length** overrides. 10s remains the **video** unit, not script length.
**Episode delivery lock:** **Standard episode ≈ 60s = 6×10s units** (boundary flex **5–7** only to land on Scene Exit / Therefore). **Finale episode(s):** **no 60s minimum**—length = remaining screenplay units only (**1–5 units OK**); **ban** pad/filler to hit 6. Group after full index; render **one unit per turn** until the master is fully produced.
**Worldview lock:** Lock **World bible** before scenes (period, tech/social rules, causality). Every beat must be reasonable **inside** that world—no anachronisms or unearned facts.
**Production style lock:** After Source research, lock **`## Production style lock`** for the **entire production** before `## Screenplay`. All refs, every 10s unit, and every ref2v call inherit this look—**no re-picking style per board, character, or scene** unless the user explicitly requests a look change.
**Unique flow:** **Language & Region Lock** (Step 0) → **FIRST_PITCH_3** when thin first input (once, 3 style+story buttons) → understand user material → localized `search_web` **×5** → deliverable **`source_notes`** → **`production_style_lock`** + **`world_bible`** → locked **`screenplay`** → Story plan + Beat budget → full **`cut_index_10s`** + **`episode_map`** + **`shot_sheet`** → refs (`gpt-image2`) → **this-turn** `read_skill_file('references/ref_prompt_guide.md')` + `read_skill_file('references/prompt_compiler.md')` + **`read_skill_file('references/continuity_state.md')`** + **`read_skill_file('references/planning.md')`** → six-part prompt → one `generate_video(model='rm3.5-ref')`. *(Anchor slugs — user-visible H2 text from **DELIVERABLE_HEADINGS[`language.code`]**, e.g. ja →