One visual family
Shared materials, proportions and studio lighting.
Official product guardrails and reusable visual-style prompts for creating personal AI assistant avatars that still feel at home in Tanka.






不决定具体画风,而是统一品牌气质、构图、安全区域和默认行为。用户没有填写风格时,自动叠加 Soft 3D。
OFFICIAL GUARDRAIL: Create one subject for an AI assistant avatar. Deliver one 1254×1254 px square PNG master for both mobile and web; product surfaces may downscale it proportionally but must not upscale it. Use a centered, front-facing bust composition. Keep the complete hairstyle, ears, glasses and identity-critical details safely inside an automatic circular crop. Target an 11% top gap above the hairstyle or ears—about 139 px in the 1254 px master; accept 10–12%, approximately 125–150 px, equivalent to about 8 px in a 72×72 px circular avatar. The overall presence should feel calm, airy, modern, trustworthy and friendly. Use pure white by default; change the background only when the user explicitly specifies one, and keep it simple. Blue may appear as a restrained environmental or wardrobe accent, but never as a cast over skin, fur, hair or clothing. No text, logo, watermark, complex scene, extra character or unrelated prop. If the user does not specify a visual style, apply the Soft 3D style module by default.四条接近日常表达的输入。未写画风时,官方护栏会自动选择 Soft 3D,并补齐白底、构图和裁剪规则。




Natural commercial portrait photography with believable detail and restrained editorial polish. Professional, approachable and never over-retouched.
The official guardrail and section structure are ready. The style-specific recipe and visual anchors are still being calibrated.
The current production-ready Skill: friendly toy-like form, controlled detail and a stable visual family across people and characters.
STYLE — SOFT 3D: Apply the official guardrail. Render a polished soft 3D toy/clay AI assistant avatar with broad sculpted forms, smooth matte-to-satin skin or fur, restrained tactile clothing texture, glossy black oval eyes, short eyebrows, a small rounded nose and one short, thin, softly recessed, gently upturned closed-mouth line close beneath the nose. Deliver four separate 1254×1254 px square PNG masters, never a collage. Use a front-facing centered bust composition designed for automatic circular cropping. Keep the complete hairstyle, ears, glasses and identity-critical details inside the circle-safe area. Target an 11% top gap above the hairstyle or ears—about 139 px, with a 125–150 px tolerance; this becomes about 8 px at 72×72 px. Use a uniform pure-white background unless the user explicitly specifies another simple background. Keep the visual family, camera, lighting, material, eye design and mouth-line language consistent across all four candidates.Shared materials, proportions and studio lighting.
Short, thin, recessed and close beneath the nose.
Hair, ears and accessories stay inside the crop.
Another background only when the user requests it.
Female, male, Shiba and owl references travel with the behavior contract, prompt recipes and quality gates.
soft-3d-avatar-style-imagegen/
├── SKILL.md
├── agents/openai.yaml
├── assets/
│ ├── style-anchor-female-blonde.png
│ ├── style-anchor-female-curly.png
│ ├── style-anchor-male-bearded.png
│ ├── style-anchor-male-navy.png
│ ├── style-anchor-animal-shiba.png
│ └── style-anchor-animal-owl.png
└── references/
├── official-guardrail.md
├── style-system.md
├── prompt-recipes.md
└── examples.mdContemporary editorial anime with clean linework, adult proportions, restrained color and soft dimensional light—not chibi, noisy or fan-art coded.
The official guardrail and section structure are ready. The style-specific recipe and visual anchors are still being calibrated.
Add the package to Codex, then describe the assistant in natural language. If no style is named, Soft 3D remains the default.
Download ZIPunzip soft-3d-avatar-style-imagegen.zip -d ~/.codex/skillsRestart Codex after unzipping so the Skill is discovered.