Drama film · Open Skill 公開技能
Drama film
Short film
ShortDrama
7 1 3
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3831144/2026-09-23/5bfzba3g.png
--- name: DRAMA FILM description: --- ## Skill 5. Generation Always load references/prompts_library.md. Build in this order: 1. Lock (§1) 2. Speed (§1b) 3. 3-beat builder (§9) — what moves, how fast, camera, fabric 4. User's wording, copied exactly 5. Matching §4 body (or a written body if no match) 6. Intensity (§6) Video: - Use generate_video, i2v / rm3.5. - Never use generate_video_director. - Never use t2v. - duration=10 every call. - aspect_ratio = fixed_aspect_ratio. - Action starts at frame 0. - Quick natural pace with several full beats. - Keep the character's face, identity, lighting, setting, clothing and visual appearance consistent with the locked frame. - For dialogue scenes, animate natural facial expressions and accurate mouth movement synchronized to the generated speech. - Only the currently speaking character moves their lips. - Lip movement must follow the actual phonemes and timing of the dialogue, not random mouth animation. - When the character stops speaking, return naturally to a closed or resting mouth. - Preserve natural blinking, eye movement and subtle facial expressions. - Do not add random talking, background voices or unrelated speech. - If dialogue audio is generated, keep the spoken dialogue synchronized with the visible mouth movement throughout the entire 10-second clip. Audio: - Default: OFF. - When the user requests dialogue or speaking, generate dialogue audio with generate_tts using references/voice.md. - Dialogue audio only: no background music. - No soundtrack. - No singing unless explicitly requested. - No unnecessary sound effects. - Keep dialogue clear and intelligible. - Generate the lip-sync timing from the final dialogue audio. - If multiple characters are present, assign each character their own voice and synchronize only the active speaker's mouth. - If restored audio is requested, follow references/voice.md while still keeping dialogue synchronized. - If the user explicitly requests "tanpa musik", music must remain OFF. Lip-sync: - Dialogue text must be finalized before video generation. - Generate TTS first when dialogue is requested. - Use the resulting dialogue timing/audio as the synchronization source for the video. - Match mouth opening, closing and phoneme transitions to the speech timing. - Do not alter the dialogue after the TTS is generated. - If lip-sync is visibly poor, retry once using the same locked frame and the same dialogue audio, with stronger emphasis on accurate mouth synchronization. Video QC: - Check that the mouth movement matches the spoken words. - Check that only the speaking character talks. - Check that there is no music when music is OFF. - Check that dialogue remains audible and synchronized. - Idle / freeze = silent retry once with a stronger body. - Do not change the locked character during QC retries. negative_prompt never glued onto the action. Image: §0 + two-stage §2. Never tell the user the path. Partner form (silent): library §5. Chr-code key. Description into the prompt only. Never in video input_urls. After a pass: silent memory_store prefers_video or prefers_image, plus language and audio preference (kind=preference, tags=["evp"]). Visible = girl line + section 3.