anima-prompt-pool
Build external prompt pools for Anima-based source generation. Use when creating many Anima prompts for later batch generation, especially when prompts must be saved in a stable JSON-friendly format with ids, categories, notes, subject tags, and natural-language scene text.
- License 许可
- MIT
- 最近更新
- 2026-10-07
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- anima-prompt-pool
- description
- Build external prompt pools for Anima-based source generation. Use when creating many Anima prompts for later batch generation, especially when prompts must be saved in a stable JSON-friendly format with ids, categories, notes, subject tags, and natural-language scene text.
02 资源文件 1
- 参考文档agents/openai.yaml191 B
03 SKILL.md 正文
Anima Prompt Pool
Overview
Create prompt-bank entries that are easy to append into the batch pipeline later. Write prompts for Anima source generation, not for the Illustrious teacher pass.
Core Rules
- Use
subject_tags + scene_promptas the canonical format. subject_tagsholds visible anchors in tag form.scene_promptholds the main natural-language scene description.- Optimize for batch consistency, not literary flair.
- Keep one entry focused on one scene idea.
- If an outfit, lighting, or camera angle matters, state it explicitly.
Output Format
Default to this JSON object format:
{
"id": "p001",
"category": ["office", "professional", "glasses"],
"exposure": "clothed",
"subject_tags": "1girl, solo, dark navy hair, long hair, teal eyes, pale skin, thin silver-rim glasses, white research blouse with rolled-up sleeves, dark navy vest, black pencil skirt, black tights, sitting at desk, legs crossed, soft skin, cool blue tones, monitor glow, upper body",
"scene_prompt": "An adult woman with long dark navy hair and teal eyes is sitting at a sleek modern office desk, wearing a white research blouse with rolled-up sleeves over a dark navy vest and black pencil skirt with black tights. She has thin silver-rimmed glasses and is focusing on a holographic display. The scene is lit by cool blue screen glow and soft ambient office lights. The image should feel professional and focused, with sharp suit textures and polished glass surfaces.",
"negative_prompt": "multiple girls, 2girls, extra person, crowd, people in background",
"notes": "Optional short note about intent or risk."
}
exposure field
exposure classifies the scene's nudity level. This drives which body/exposure tags MUST appear in subject_tags and ensures caption–image consistency for LoRA training.
| exposure | Definition | Required tags (subject_tags) | Forbidden tags (subject_tags) |
|---|---|---|---|
clothed |
Fully dressed, no sexual parts visible | Explicit clothing items (white blouse, black skirt, etc.), clothed |
nude, nipples, pussy, bare_breasts, topless |
partial |
Partially exposed — underwear visible, cleavage, sideboob, lifted skirt, see-through, etc. | Specific exposure state (cleavage, black lace bra, unbuttoned shirt, sideboob, lifted skirt, panties visible, etc.) |
nude, fully_clothed |
nude |
Fully nude or near-nude, sexual parts directly visible | nude or topless, plus visible anatomy (nipples, navel, bare_breasts, pussy, etc. as applicable to the composition) |
clothed, covered_nipples, fully_clothed |
Rules:
- Every prompt MUST have an
exposurefield. No default — force explicit classification. - The tags in
subject_tagsMUST be consistent with the declaredexposure. Ifexposureisclothed, do not includenipples. Ifexposureisnude, do not includeclothed. - For
partial, specify exactly what is exposed and what remains covered. Ambiguity here causes the worst training noise. exposurevalue is also carried intocategoryautomatically for filtering (e.g.,["bedroom", "nude"]).
negative_prompt field
negative_promptis optional per-item. If omitted, the pipeline uses the global default frompipeline_meta.json(source_default_negative/teacher_default_negative).- Use per-item negatives only when a scene needs extra suppression beyond the global default (e.g., a crowded street scene that tends to spawn extra characters, or a mirror scene that generates duplicates).
- Common per-item negatives:
- Crowded environments:
multiple girls, 2girls, extra person, crowd, people in background - Mirror/reflection scenes:
duplicate, clone, symmetry, reflection - Indoor scenes with windows:
extra person, crowd, people in background
- Crowded environments:
- Do not duplicate quality negatives (
worst quality, low quality) here — those are already in the global default. - For multi-character scenes (2girls etc.), remove the anti-multi tags from negative:
"negative_prompt": ""to override the global default.
When asked for many prompts, output a JSON array of those objects.
Legacy compatibility:
- old
promptsingle-field format may still exist - but new pools should default to
subject_tags + scene_prompt
Real File Example
Use a prompt-bank.json file like this:
[
{
"id": "p001",
"category": ["office", "professional", "glasses", "clothed"],
"exposure": "clothed",
"subject_tags": "1girl, solo, dark navy hair, long hair, teal eyes, pale skin, thin silver-rim glasses, white research blouse with rolled-up sleeves, dark navy vest, black pencil skirt, black tights, sitting at desk, legs crossed, clothed, soft skin, cool blue tones, monitor glow, upper body",
"scene_prompt": "An adult woman with long dark navy hair and teal eyes is sitting at a sleek modern office desk, wearing a white research blouse with rolled-up sleeves layered under a dark navy vest, paired with a black pencil skirt and black tights. She wears thin silver-rimmed glasses. She is focusing on a holographic display. The scene is lit by cool blue screen glow and soft ambient office lights. The image should feel professional and focused, with sharp suit textures and polished glass surfaces.",
"notes": "Good for polished interior surfaces."
},
{
"id": "p002",
"category": ["nature", "forest", "ethereal", "clothed"],
"exposure": "clothed",
"subject_tags": "1girl, solo, dark navy hair, long hair, teal eyes, pale skin, white cotton sundress, thin shoulder straps, barefoot, walking, clothed, soft skin, warm tones, dappled sunlight, full body",
"scene_prompt": "An adult woman with long dark navy hair and teal eyes is walking through a sun-drenched forest, wearing a white cotton sundress with thin shoulder straps. She is barefoot on the mossy ground. The scene is lit by warm dapples of sunlight filtering through the canopy of ancient mossy trees with floating pollen. The image should feel ethereal and serene, with soft fabric movement and detailed leaf textures.",
"notes": "Good for soft natural light."
},
{
"id": "p003",
"category": ["bedroom", "intimate", "partial"],
"exposure": "partial",
"subject_tags": "1girl, solo, dark navy hair, long hair, teal eyes, pale skin, unbuttoned white dress shirt, black lace bra visible, cleavage, bare legs, sitting on bed edge, leaning back on hands, soft skin, warm tones, dim lamp light, full body",
"scene_prompt": "An adult woman with long dark navy hair and teal eyes is sitting on the edge of a bed, wearing an unbuttoned white dress shirt that reveals a black lace bra underneath. Her legs are bare. She is leaning back on her hands with a relaxed expression. The scene is lit by warm dim lamp light from a bedside table. The image should feel intimate and natural, with soft fabric folds and warm skin tones.",
"notes": "Partial exposure — bra visible through unbuttoned shirt."
},
{
"id": "p004",
"category": ["bathroom", "morning", "nude"],
"exposure": "nude",
"subject_tags": "1girl, solo, dark navy hair, long hair, wet hair, teal eyes, pale skin, nude, bare breasts, nipples, navel, standing, holding towel at side, wet skin, water droplets, soft skin, warm tones, soft morning light, full body",
"scene_prompt": "An adult woman with long dark navy hair (wet and clinging to her shoulders) and teal eyes is standing nude in a steamy bathroom, holding a white towel casually at her side. Water droplets glisten on her pale skin. The scene is lit by soft warm morning light coming through a frosted window. The image should feel natural and candid, with realistic water droplet textures and steam diffusion.",
"notes": "Full nude — all relevant anatomy tags included."
}
]
Use a review-queue.json file like this:
[
{
"id": "p001",
"status": "pending",
"source_image": "images/source/p001.png",
"teacher_image": "images/teacher/p001.png",
"subject_tags": "1girl, solo, long navy hair, silver-rimmed glasses, white research blouse, dark navy vest, black pleated skirt",
"scene_prompt": "An adult woman is sitting at a sleek modern office desk, focusing on a holographic display. The scene is lit by cool blue screen glow and soft ambient office lights. The image should feel professional and focused, with sharp suit textures and polished glass surfaces.",
"category": ["office", "professional", "glasses"],
"notes": "Good for polished interior surfaces."
}
]
Recommended status values:
pending— not reviewed yetyes— approved for datasetno— rejectedhold— keep for later review
Prompt Construction Workflow
For each prompt:
-
Choose one clear scene family
- neon city
- luxury interior
- futuristic stage
- bedroom
- bathroom
- office
- fantasy tavern
-
Build
subject_tagswith all six axes filled subject_tags is used BOTH for Anima source generation AND as the Illustrious teacher prompt (without natural language). It must be rich enough for the teacher to render the scene correctly on its own.Required axes (always include all six):
Axis Examples Why Character identity 1girl, solo, dark navy hair, long hair, teal eyes, pale skinTeacher doesn't know the character otherwise Outfit / exposure white sundress, barefoot, clothedornude, bare breasts, nipplesorunbuttoned shirt, black lace bra visible, cleavage— must match theexposurefieldAmbiguous nudity level causes rendering errors AND training noise Pose / anatomy standing, arms at sidesoron back, lying on bed, knees up, navelPrevents body distortion, especially in lying/crouching poses Skin / texture state soft skinorsweating, wet skinoroiled skin, sweat dropletsDirectly affects erotic scene quality Color / lighting tone warm tones, golden lightorcool blue tones, dim lightingorhigh contrast, neon glowAnchors the teacher's color grading direction Composition upper bodyorfull bodyorclose-up, face focusPrevents framing mismatch between source and teacher Example — bedroom nude (exposure: "nude"):
1girl, solo, dark navy hair, long hair, teal eyes, pale skin, nude, bare breasts, nipples, navel, flat stomach, on back, lying on bed, arms above head, soft skin, messy hair, warm tones, dim lighting, full bodyExample — office clothed (exposure: "clothed"):
1girl, solo, dark navy hair, long hair, teal eyes, pale skin, thin silver-rim glasses, white research blouse with rolled-up sleeves, dark navy vest, black pencil skirt, black tights, clothed, sitting at desk, legs crossed, soft skin, cool blue tones, monitor glow, upper bodyExample — bedroom partial (exposure: "partial"):
1girl, solo, dark navy hair, long hair, teal eyes, pale skin, oversized white t-shirt, no bra, sideboob, black panties, bare legs, sitting on bed, knees up, soft skin, warm tones, morning light, full body -
Write
scene_promptas natural language with explicit visual detailsscene_prompt: action, environment, lighting, mood, material richness- This is used by Anima only (not teacher), so it can be descriptive prose
- MUST explicitly describe: hair color & style, clothing items with color/material/detail, accessories, footwear
- The first sentence should establish the character's appearance: "An adult woman with long dark navy hair and teal eyes is..."
- Clothing descriptions should be specific: not just "wearing a dress" but "wearing a white cotton sundress with thin shoulder straps"
- If exposure is
partialornude, describe the state of dress/undress naturally in the prose - Accessories (glasses, jewelry, hair ties, etc.) should be mentioned if present
-
Decide if per-item negative is needed
- Most scenes: omit
negative_prompt(global default handles it) - Busy backgrounds (streets, crowds, festivals): add
"negative_prompt": "multiple girls, 2girls, extra person, crowd, people in background" - Multi-character scenes: add
"negative_prompt": ""to clear the global anti-multi default - Mirror/window scenes: add
"negative_prompt": "duplicate, clone, reflection"
- Most scenes: omit
-
Add a short note only if needed
- good for glossy clothing
- strong color contrast
- risk of cluttered background
Writing Pattern
Use this skeleton:
exposure: [clothed | partial | nude]
subject_tags: [인원], [캐릭터 외형: 머리색, 머리길이, 눈색, 피부], [의상 또는 노출 상태 — exposure와 일치], [포즈/해부학], [피부 질감], [컬러/라이팅 톤], [구도]
scene_prompt: An adult woman with [hair color & style] and [eye color] is [pose/action] in [environment], wearing [specific clothing with color/material/detail OR nude state description]. [Accessories if any]. The scene is lit by [lighting]. The image should feel [mood], with [material/color/composition details].
negative_prompt: (optional, only if scene background is busy or multi-character suppression is needed)
Checklist before submitting a prompt — all seven items present?
- [ ]
exposurefield set (clothed / partial / nude) - [ ] Character identity (hair color & style, eyes, skin)
- [ ] Outfit or exposure tags consistent with
exposurefield - [ ] Pose and anatomy anchors
- [ ] Skin/texture state
- [ ] Color/lighting tone
- [ ] Composition (framing)
- [ ] scene_prompt explicitly describes hair, clothing/nudity, and accessories in natural language
Constraints
- Prefer prompts that remain readable after batch generation.
- Do not overload every prompt with too many visual ideas.
- Do not write teacher-stage trigger tags here unless explicitly requested.
- Do not convert the whole entry into tags only; keep
scene_promptin natural language. - If the user wants erotic prompts, keep the same structured format and focus on scene clarity.
If the User Requests Variations
Generate variations by changing only one or two axes:
- environment
- lighting
- outfit material
- color palette
- camera distance
Do not change every axis at once unless the user asks for maximum variety.
Deliverables
When asked to prepare a pool, provide:
- A JSON array
- Optional category summary
- Optional warnings about prompts likely to fail in batch
04 同仓库更多 Skills
generate-video
claude-play-bridge MCP 서버로 영상을 생성한다. 로컬 ComfyUI의 영상 워크플로 패키지(MiniMax H3 = 영상+오디오 동시 생성, WAN 2.2 i2v, Z-Image→영상)를 호출한다. 렌더가 수 분~수십 분 걸리므로 async 큐잉 + 파일 폴링이 기본 절차다. "영상", "동영상", "비디오", "video", "i2v", "t2v", "클립 만들어", "움직이는", "애니메이션 영상" 요청에 트리거된다.
generate-image
claude-play-bridge MCP 서버로 이미지를 생성한다. 기본 백엔드는 ComfyUI(로컬 GPU, 워크플로 패키지 기반)이고, 사용자가 명시적으로 요청할 때만 GPT/Gemini 백엔드를 사용한다.
generate-image-openai
OpenAI/GPT 이미지 생성(GPT Image 품질 / 텍스트 렌더링·이미지 편집 강점). 기본 백엔드는 Codex CLI 구독(건당 과금 없음). ⚠️ 사용자가 이 백엔드를 명시적으로 지정했을 때만 사용한다(예 "GPT로/OpenAI로 그려줘"). 기본 이미지 생성은 generate-image(ComfyUI)이며, 사용자가 콕 집어 요청하지 않았다면 절대 OpenAI를 자동으로 선택하지 마라.
generate-spritesheet
Gemini 이미지 생성으로 게임용 아이콘 스프라이트시트를 만들고 자동 슬라이스한다. 아이콘, UI 리소스, 뱃지, 마커 등 여러 개의 동일 카테고리 에셋이 필요할 때 사용한다.
generate-image-gemini
Gemini API로 이미지를 생성한다. 패널 배경, UI 리소스, 아이콘, 장면 삽화 등 디자인 리소스 제작에 사용한다. ComfyUI가 없을 때의 이미지 생성 대안이기도 하다.
workflow-research
새 ComfyUI 워크플로우를 설계하거나 신규 모델, 양자화 모델, LoRA 개발 전략을 비교할 때 분석 틀을 잡는다.