## 它是什么

MOSS-Image 的提示词编写规范（纯文本 skill，不含任何模型权重）。

把一句中文需求编译成一段可直接投喂 MOSS-Image 的英文正向提示词，按画面复杂度自动选择

三种输出形态之一：连贯自然语言、闭合标签分块、Layout JSON。

适用于 MOSS-image模型

## 它解决什么问题

MOSS 用 Qwen3.5-9B 做文本编码，对空间关系敏感。写提示词最常见的四种崩法：

人物悬空没重心、交叠的肢体分不清归属、深度与体液描述互相矛盾、写了机位根本看不见的部位。

这份规范把四类问题拆成可执行的检查：支撑三元组、动作阶段锁定、可见性三步、孔位锁定。

## 输出契约

- 默认只输出一段正向提示词；负向词与采样参数仅在明确要求时另列

- 自然语言里不出现坐标，坐标只进 Layout

- 正文默认英文，图内文字语言独立决定

- 无标题、无分析、无中间草稿

## 包含什么

15 节，736 行：

- 输入处理：自然语言 / Danbooru tags / 参考图三种输入

- 形态路由：四问判定，格式只路由一次

- 风格推导：按需现场推导，无预设档位、无「名字→固定句」对照表

- 人物控制：骨相六层组合（年龄事实 → 头脸比例 → 轮廓五官 → 颈肩框架 → 软组织 → 体态）

- 多人隔离：方位分块 + 反向阻断防串脸串灯

- 图内文字与界面：气泡形状推导、HUD 部件写法、修复纪律

- Layout 工程：草稿 schema、中心原点 yxyx 换算、多分镜母版

- 接触与成人几何：八层检查表、三级触发、动作阶段、可见性推导、体液量级校准

- 成品验收三组：结构完整性 S1–S7、逐句映射、事实与格式 F1–F7 / G1–G7

## 适用

- 用 MOSS 出写实人像、生活场景、多人群像

- 需要精确排版的海报 / UI / 多分镜漫画（走 Layout JSON）

- 成年人之间的亲密与接触场景（空间关系复杂，这份规范主要针对这类画面打磨）

## 边界

仅处理成年人之间的接触与性内容。涉及未成年人的性内容不产出——明示年龄、外观指向、

设定指向三条任一命中即停止。这是全文唯一的内容限制，其余题材（年龄、人数、体位、

体型、露骨程度）全部按使用者指定写，不做换算。

## 用法

把 SKILL.md 放进你所用客户端的技能目录，或在会话里直接作为系统提示加载。

之后给需求即可，输出就是成品提示词，无需二次整理。

## 来源与致谢

本规范不是原创方法论，是基于两份公开资料整理改写而成：

- 基线：[Moeblack/moss-image-prompt-skill](https://github.com/Moeblack/moss-image-prompt-skill)

取其 moss_fusion_spec_v2.md 作为结构底本，保留形态路由、分块纪律、骨相体系与 Layout 工程。

- 增量：[Moody Moss Prompt Writing Skill (LLM System Prompt)](https://civarchive.com/models/3010146/moody-moss-prompt-writing-skill-llm-system-prompt)

取其成人几何部分的四条硬规则：孔位锁定、作物门控、体内不可见、体液量级校准。

改写部分：把原稿的固定风格档位改为按需推导（无预设、无「名字→固定句」对照表）；

章节从 20 节合并到 15 节，去掉重复的验收清单与媒介专节；

成品验收统一成三组（结构完整性 S1–S7 / 逐句映射 / 事实与格式 F1–F7、G1–G7）；

加入成年人内容边界，其余内容尺度按使用者指定。

本文件为独立改写，非上述两处的官方发布物，不含任何模型权重，与原作者无隶属关系。

原始资料的著作权归各自作者，使用与再分发请遵循原仓库与原页面的许可条款。

## What it is

A prompt-writing specification for MOSS-Image (plain-text skill, no model weights included).

It compiles a short request into one ready-to-paste English positive prompt, and picks one

of three output forms based on how complex the scene is: continuous natural language,

closed-tag blocks, or Layout JSON.

Works with MOSS-Image models.

## What problem it solves

MOSS uses Qwen3.5-9B as its text encoder and is sensitive to spatial relationships. The four

most common failures when writing prompts for it:

a figure floating with no weight bearing, overlapping limbs with no clear ownership, depth and

fluid descriptions contradicting each other, and describing parts the camera cannot see.

This specification turns all four into executable checks: the support triad, the action-stage

lock, three-step visibility derivation, and orifice locking.

## Output contract

- One positive prompt by default; negative prompt and sampling parameters only when explicitly asked

- No coordinates in natural language; coordinates appear in Layout only

- Body text defaults to English; in-image text language is decided independently

- No headings, no analysis, no intermediate drafts

## What it contains

15 sections, 736 lines:

- Input handling: natural language / Danbooru tags / reference images

- Form routing: four-question decision, routed exactly once

- Style derivation: derived on the spot from the request, no preset tiers, no "name → fixed sentence" lookup table

- Character control: six-layer bone-structure assembly (age fact → head-face proportions → contour and features → neck-shoulder frame → soft tissue → build)

- Multi-character separation: positional blocks plus reverse blocking to stop feature and lighting bleed

- In-image text and UI: bubble-shape derivation, HUD component phrasing, repair discipline

- Layout engineering: draft schema, central-origin yxyx conversion, multi-panel master template

- Contact and adult geometry: eight-layer checklist, three-level trigger, action stage, visibility derivation, fluid quantity calibration

- Three-part output check: structural completeness S1–S7, per-sentence mapping, facts and format F1–F7 / G1–G7

## Use cases

- Photorealistic portraits, everyday scenes, and group shots on MOSS

- Poster / UI / multi-panel comic layouts that need exact placement (via Layout JSON)

- Intimate and contact scenes between adults (spatially demanding; this specification was mainly refined against such frames)

## Boundary

Handles contact and sexual content between adults only. Sexual content involving minors is not

produced — any one of three triggers (stated age, appearance suggesting a minor, setting

suggesting a minor) stops it. This is the only content limit in the document; every other

element (age, number of people, position, body type, explicitness) is written exactly as the

user specifies, with no adjustment.

## Usage

Drop SKILL.md into your client's skill directory, or load it directly as a system prompt in the

session. After that, give it the request — the output is a finished prompt with no cleanup needed.

## Credits

This is not an original methodology. It is an adaptation of two public sources:

- Baseline: [Moeblack/moss-image-prompt-skill](https://github.com/Moeblack/moss-image-prompt-skill)

Used moss_fusion_spec_v2.md as the structural base, keeping form routing, block discipline,

the bone-structure system and Layout engineering.

- Additions: [Moody Moss Prompt Writing Skill (LLM System Prompt)](https://civarchive.com/models/3010146/moody-moss-prompt-writing-skill-llm-system-prompt)

Took four hard rules from its adult-geometry section: orifice locking, crop gating,

internal invisibility, and fluid quantity calibration.

What changed: fixed style tiers replaced with on-demand derivation (no presets, no "name → fixed

sentence" table); sections merged from 20 down to 15, with duplicated checklists and per-medium

chapters removed; output checks consolidated into three groups (structure S1–S7 / per-sentence

mapping / facts and format F1–F7, G1–G7); an adults-only boundary added, with every other content

parameter written as the user specifies.

This is an independent adaptation. It is not an official release of either source, contains no

model weights, and is not affiliated with the original authors. Copyright in the source material

remains with its respective authors; reuse and redistribution follow the licences of the original

repository and page.