Imported from devlitus/space-flight---combat-simulator (
.codex/skills/character-generator/SKILL.md). Install upstream withnpx skills add devlitus/space-flight---combat-simulator --skill character-generator. Copyright stays with the author.
The skill creates:
A Character Reference Sheet. Individual images for each character pose. Optional facial expression assets. Production-ready character assets that can later be used for AI video generation.
The primary goal is to maintain visual character consistency across multiple generated images.
Objective
Given a character description, generate a reusable character asset package containing the same character in multiple poses.
The character must maintain:
Visual identity. Facial features. Hairstyle. Skin tone. Clothing. Accessories. Body proportions. Apparent age. Artistic style. Visual palette. Level of detail.
The pose may change, but the character's identity must remain consistent.
Input
The skill receives a CharacterDefinition object.
{ "name": "Alex", "age": 28, "gender": "male", "appearance": { "hair": "short dark brown hair", "eyes": "green", "skin": "light skin", "body": "athletic" }, "clothing": { "top": "dark blue jacket", "bottom": "black jeans", "shoes": "white sneakers" }, "accessories": [ "black watch" ], "personality": [ "confident", "friendly" ], "art_style": "cinematic realistic", "poses": [ "standing", "walking", "running", "sitting", "waving", "jumping" ] } Output
The skill should produce an asset structure similar to:
characters/ └── alex/ ├── reference/ │ └── character-reference.png │ ├── poses/ │ ├── standing.png │ ├── walking.png │ ├── running.png │ ├── sitting.png │ ├── waving.png │ └── jumping.png │ └── metadata.json Resolution Character Reference Sheet
Preferred resolution:
2048 × 2048 px
Aspect ratio:
1:1
The character should be shown full-body.
Individual Pose Assets
Each pose should preferably be generated at:
2048 × 2048 px
Aspect ratio:
1:1
The character should occupy approximately 70–90% of the image.
Leave enough space around the character to prevent cropping.
Phase 1 — Character Reference Sheet
Before generating individual poses, create a character reference sheet.
The reference should contain:
Front view. 3/4 view. Side view. Back view. Full body. Neutral facial expression. Complete clothing. All important accessories.
Example:
FRONT 3/4 SIDE BACK
O O O O
/|\ /|\ /|\ /|\
/ \ / \ / \ / \
The reference image should use a clean, neutral background.
Avoid complex environments.
Do not include objects that could be confused with parts of the character.
Phase 2 — Identity Lock
After generating the reference sheet, create a structured description of the character.
Example:
{ "identity": { "face": "oval face", "eyes": "green almond-shaped eyes", "hair": "short dark brown hair", "nose": "medium straight nose", "skin": "light skin" },
"body": { "height": "tall", "build": "athletic", "proportions": "realistic" },
"clothing": { "jacket": "dark blue jacket", "pants": "black jeans", "shoes": "white sneakers" },
"accessories": [ "black watch" ] }
This information should be treated as immutable during pose generation.
Phase 3 — Pose Generation
Generate one independent image for each pose.
Each generation should use:
Character Reference + Identity Description + Pose Description + Visual Style
Never rely exclusively on text to reconstruct the character.
Pose Prompt Template
Use the following structure:
Generate a full-body image of the exact same character shown in the reference image.
CHARACTER IDENTITY: {identity_description}
CLOTHING: {clothing_description}
STYLE: {art_style}
POSE: {pose_description}
CAMERA: Full-body shot, eye-level camera, centered composition.
COMPOSITION: The entire body must be visible from head to feet. No cropping. Natural anatomy. Natural body proportions.
CONSISTENCY: Preserve the exact same:
- face
- hairstyle
- eye color
- skin tone
- body proportions
- clothing
- shoes
- accessories
- age
- visual style
Only change the character's pose.
BACKGROUND: Simple neutral background.
OUTPUT: High-resolution, clean silhouette, production-ready character asset. Supported Poses
The skill should support at least:
standing walking running sitting lying jumping waving pointing looking_back arms_crossed hands_in_pockets thinking talking laughing angry surprised scared
The skill should also support custom poses.
Example:
{ "pose": "The character is running toward the camera while looking over his shoulder." } Pose Rules
Each generated pose must:
Maintain realistic anatomy. Maintain consistent body proportions. Maintain identical clothing. Maintain all accessories. Maintain recognizable facial features. Avoid distorted hands. Avoid extra fingers. Avoid duplicated limbs. Avoid clothing changes. Avoid age changes. Avoid style changes.
The pose should be the primary visual element that changes.
Negative Constraints
When the image generation model supports negative prompts, use:
different character, different face, different hairstyle, different clothing, different shoes, different body proportions, different age, deformed anatomy, extra arms, extra legs, extra fingers, missing fingers, duplicate limbs, cropped body, cut off feet, cut off head, distorted face, blurred face, low resolution, text, logo, watermark, busy background Character Consistency
Character consistency has priority over visual variation.
If there is a conflict:
Character Identity > Pose > Composition > Background
Never modify the character's identity to achieve a requested pose.
Background
Individual character assets should preferably use:
plain neutral background
or:
transparent background
when supported correctly by the generation model.
Do not initially generate the character inside complex environments.
Environments should be generated as independent assets.
Asset Separation
The skill should treat the following as independent assets:
Character Pose Expression Environment Props Camera Lighting
Do not combine all these elements into a single image during the initial character creation process.
Example:
Character + Pose + Environment + Camera = Scene
This allows the character to be reused across multiple scenes.
Expressions
Facial expressions should preferably be generated independently from poses.
Example:
expressions/ ├── neutral.png ├── happy.png ├── sad.png ├── angry.png ├── surprised.png └── scared.png
Expressions must not modify:
Facial identity. Apparent age. Hairstyle. Body proportions. Character identity. Metadata
Create a metadata.json file.
Example:
{ "character_id": "alex", "name": "Alex", "version": "1.0", "reference_image": "reference/character-reference.png",
"identity": { "age": 28, "hair": "short dark brown hair", "eyes": "green", "skin": "light", "body": "athletic" },
"clothing": { "top": "dark blue jacket", "bottom": "black jeans", "shoes": "white sneakers" },
"poses": [ "standing", "walking", "running", "sitting", "waving", "jumping" ],
"resolution": { "width": 2048, "height": 2048 } } Generation Workflow
The agent must follow this workflow:
USER │ ▼ Character Definition │ ▼ Generate Character Reference │ ▼ Validate Identity │ ├── FAIL ──► Regenerate Reference │ ▼ Lock Character Identity │ ▼ Generate Pose 01 │ ▼ Validate Consistency │ ├── FAIL ──► Regenerate Pose │ ▼ Generate Pose 02 │ ▼ Validate Consistency │ ├── FAIL ──► Regenerate Pose │ ▼ ... │ ▼ Generate Metadata │ ▼ Character Asset Package Validation
After every generated image, validate the following.
Identity [ ] Same face [ ] Same hairstyle [ ] Same eye color [ ] Same skin tone [ ] Same apparent age Body [ ] Same proportions [ ] Same body type [ ] Correct anatomy [ ] No duplicated limbs Clothing [ ] Same clothes [ ] Same colors [ ] Same shoes [ ] Same accessories Pose [ ] Requested pose is correct [ ] Full body is visible [ ] No cropping Quality [ ] High resolution [ ] Sharp image [ ] Clean silhouette [ ] No text [ ] No watermark Important Rule
Do not generate all poses as a single image when the goal is to create reusable animation assets.
Use:
Character Reference Sheet │ ├── Pose 01 ├── Pose 02 ├── Pose 03 ├── Pose 04 ├── Pose 05 └── Pose N
The Character Reference Sheet is used to maintain identity.
The individual pose images are the reusable production assets.
Recommended Asset Architecture
For an AI movie or video generation system:
PROJECT │ ├── characters/ │ └── alex/ │ ├── reference/ │ │ └── character-reference.png │ │ │ ├── poses/ │ │ ├── standing.png │ │ ├── walking.png │ │ ├── running.png │ │ └── jumping.png │ │ │ ├── expressions/ │ │ ├── neutral.png │ │ ├── happy.png │ │ └── angry.png │ │ │ ├── wardrobe/ │ │ │ └── metadata.json │ ├── environments/ │ ├── city/ │ ├── office/ │ └── forest/ │ ├── props/ │ ├── car/ │ ├── phone/ │ └── laptop/ │ └── scenes/ ├── scene-001/ ├── scene-002/ └── scene-003/ Agent Instructions
The agent must treat the character as a reusable production asset, not as an isolated image.
Each character must have persistent identity.
Before generating a new image:
Retrieve the character reference. Retrieve metadata.json. Retrieve the locked character attributes. Apply only the requested changes. Generate the new pose. Validate consistency. Save the asset using a descriptive filename.
Never reconstruct the character from scratch when a reference already exists.
Consistency Reviewer
The generation pipeline should include a dedicated Consistency Reviewer agent.
Its responsibility is to compare every generated pose against the Character Reference.
Example:
Character Reference │ ▼ Pose Generator │ ▼ Generated Pose │ ▼ Consistency Reviewer │ ├── PASS ──► Save Asset │ └── FAIL ──► Regenerate Pose
The reviewer should evaluate:
Identity similarity Face similarity Hair consistency Clothing consistency Body proportions Color consistency Accessory consistency Anatomical correctness Pose correctness
The reviewer should return:
{ "status": "PASS", "score": 0.94, "identity_consistency": 0.97, "clothing_consistency": 0.95, "pose_accuracy": 0.91, "issues": [] }
If the score is below the configured threshold:
status = FAIL
the pose must be regenerated.
Success Criteria
The skill is successful when:
✓ Character Reference exists ✓ One independent image exists per pose ✓ All images maintain character identity ✓ All images maintain clothing ✓ All images maintain body proportions ✓ Poses are clearly different ✓ Full body is visible ✓ Images have sufficient resolution ✓ Assets can be reused in future scenes ✓ metadata.json exists ✓ Every pose passes consistency validation Design Principle
The core principle of this skill is:
ONE CHARACTER │ ▼ ONE IMMUTABLE IDENTITY │ ├──────────────┬──────────────┐ ▼ ▼ ▼ Pose Pose Pose │ │ │ ▼ ▼ ▼ Asset Asset Asset │ │ │ └──────────────┼──────────────┘ ▼ SCENES │ ▼ VIDEO
The character identity should remain stable while poses, expressions, environments, camera angles, lighting, and scenes can change independently.