SuperAPI AI Directory

extractor

General & other

View repository

Updated 10-15 17:33

About this skill

Character, Scene, and Prop Extraction Guide

Character Extraction Specifications

Extracted character fields (corresponding one-to-one with the save_dedup_characters tool parameters):

  • name(required): full name of the character
  • role: character role, protagonist/supporting role/extra
  • appearance: appearance description (300-500 words)—gender, sense of age, facial features, physique, temperament.Do not output the character's personality traits separately; transform them into external temperament and expressions and integrate them into the appearance description(for example, "cold personality" should be written as "cold gaze, restrained expression, seldom smiles")
  • styling: makeup and styling—hairstyle, clothing, makeup, accessories, etc.
  • description: background story and character relationships (optional supplement)

Scene Extraction Specifications

Extracted scene fields (corresponding one-to-one with the save_dedup_scenes tool parameters):

  • location(required): specific location name
  • time: time period (such as daytime/dusk/late night); the same location at different time periods is regarded as a new scene
  • prompt: scene description—space, furnishings, period texture, key visual elements (pure background, no characters)
  • lighting: scene lighting and shadow—light source, color tone, brightness and darkness, atmosphere

Prop Extraction Specifications

Core principle: better to extract fewer than to extract too many. Props are high-cost assets used to generate white-background single-item images and for video close-up references; only plot-critical props are worth extracting. An episode usually has 0-3key props; when there are more than 3, sort by plot importance and keep only the first 3.

Mustsimultaneously satisfythe following two conditions; neither can be missing:

  1. Directly drives the plot: the appearance, handover, damage, or discovery of this item will trigger a plot turning point (such as a murder weapon, a token, key documents, a love token, key evidence).
  2. Worth generating an image separately: subsequent storyboards will give it close-up shots or it will appear repeatedly, requiring a fixed appearance.

Three Judgment Questions(ask and answer each candidate prop yourself; if any answer is "no," abandon it):

  • ① If it is deleted, does the plot still hold? → If yes, thendo not extract(it is merely a prop-style background element)
  • ② Is it just an everyday item casually used by a character (phone, chopsticks, water cup, cigarette, umbrella)? → If yes, thendo not extract
  • ③ Is it part of the scene furnishings (tables and chairs, lamps, doors and windows, hanging paintings, tableware)? → If yes, thendo not extract(these belong to the scene description)

Typical examples that do not count as props: ordinary items used casually but that do not affect the direction of the plot; scene furnishings and furniture; items mentioned only once with no further development; a character's regular attire (classified under character makeup and styling).

If there are no props that meet the conditions,do not force extraction; when calling save_dedup_props , just pass an empty array.

Extracted prop fields (corresponding one-to-one with the save_dedup_props tool parameters):

  • name(required): prop name
  • type: type—everyday/weapon/transportation/decoration/document, etc.
  • description: Item appearance—describe only the item's own physical appearance (material, color, shape, size, condition, signs of wear, etc.); do not write about its plot purpose, and do not involve associations with characters or other things

PropsNo need to output image prompts—The final prompt for a prop is generated specifically by the prompt generation Agent before image generation (white-background single-item specification).

Usage steps

  1. Call read_script_for_extraction to read the current episode's script
  2. Call read_existing_characters to view the project's existing characters and the characters already associated with the current episode
  3. Call read_existing_scenes to view the project's existing scenes and the scenes already associated with the current episode
  4. Call read_existing_props to view the project's existing props and the props already associated with the current episode
  5. Extract only the characters, scenes, and props actually involved in the current episode
  6. Call save_dedup_characters to save characters and automatically associate them with the current episode
  7. Call save_dedup_scenes to save scenes and automatically associate them with the current episode
  8. Call save_dedup_props to save props and automatically associate them with the current episode

Current episode rules

  • The goal is to fill in the characters, scenes, and props needed for the "current episode," not to rescan the entire project
  • If something already exists in the project but is not associated with the current episode, it should still be reused and associated with the current episode
  • Deduplication rules: characters/props are matched exactly by name, scenes are matched exactly by [location + time period]; when a match is found, reuse it first and do not create duplicates
  • Near-name deduplication: when a name has a parenthetical qualifier or alias, compare by the main subject before the parentheses (e.g., "Lin Xiaoyu (protagonist)" and "Lin Xiaoyu" are treated as the same character/prop, and the existing one is reused); the normalized_name returned by read_existing_characters / read_existing_props is the normalized name, and the same applies to a scene's normalized_location; judge based on that