yashjitpal/bettergravity/hatch-pet

Hatch Pet

Create the pet the user describes using Antigravity's image-generation tool. Preserve reference art, the requested style, face, markings, proportions, and props across every animation. Infer a friendly name when one is not supplied. Pixel, plush, clay, sticker, vector, and other readable styles are supported.

Use Antigravity's native generate_image tool for visual generation, with all reference images attached. Tool names may differ between host versions: use the available native image-generation capability. If none is available, explain that limitation and stop before claiming that art was generated. Do not substitute a hand-drawn or duplicated static sprite for an AI-generated animated pet. This skill does not require a Codex account, a Codex skill, or a separate API key.

The bundled Python helpers perform deterministic processing only. Use an available Python 3 interpreter with Pillow. Check from PIL import Image; if Pillow is missing, install it in a local virtual environment for this run. Keep shell commands appropriate to the user's operating system and quote paths containing spaces.

Read the generation and assembly workflow before generating poses. Read animation rows for the exact state semantics, and the QA rubric before accepting the result.

Visible progress

Keep this four-step checklist visible in the conversation, with one step active at a time:

1. Getting the pet ready.
2. Imagining the pet's main look.
3. Picturing the pet's poses.
4. Hatching the pet.

Use the pet's name in those steps once known. Show the main character image after it exists. At the end, show the contact sheet and a working animation preview using Antigravity's normal image/artifact presentation. Use absolute local image paths. Do not replace previews with a wall of file paths or claim that an unverified image is ready.

The Pets picker also shows these stages from actual files. Its progress is written by scripts/pet_bridge.py:

python "<skill>/scripts/pet_bridge.py" begin --name "Pet Name"

Read the returned run path. Pass that exact path as --output-dir to prepare_pet_run.py, with --force to retain the new progress file. Add the user's description through --pet-notes, the chosen --pet-name, and any --reference image paths. The default library lives under the BetterGravity application-data directory; an explicit library path returned by the Create action can be supplied with the bridge's global --library option.

Update stages only after their prerequisite exists:

python "<skill>/scripts/pet_bridge.py" progress --run "<run>" --stage imagining
python "<skill>/scripts/pet_bridge.py" progress --run "<run>" --stage posing --preview references/canonical-base.png
python "<skill>/scripts/pet_bridge.py" progress --run "<run>" --stage hatching --preview references/canonical-base.png

imagining requires the prepared request; posing requires the generated canonical base; hatching requires the extracted frame manifest. If a tool fails or generation is blocked, record --stage error --message "<short explanation>" and report the actual blocker. Resume the same run after a repair.

Generation contract

• Generate one canonical base, coherent strips for all nine animation states, four cardinal look anchors, and two coherent eight-pose look strips. Every row uses the canonical base as an image reference.
• Never ask the image model to produce an entire 8 × 11 atlas. Generate row strips, then assemble them deterministically.
• Use a flat chroma key that does not occur in the pet. Backgrounds, floor shadows, floating marks, text, and guide lines must not become part of the sprite.
• The final image is a transparent PNG or WebP, exactly 1536 × 2288, arranged as 8 columns × 11 rows, with 192 × 208 cells.
• Rows 0–8 are idle, moving right, moving left, waving, jumping, failed, waiting, working, and review. Working means focused task activity, distinct from directional movement.
• Rows 9–10 are sixteen clockwise look directions. The first direction is up, followed by 22.5-degree increments; right is 90°, down 180°, and left 270°.
• Preserve approved rows during repairs. Regenerate a bad look row coherently rather than inserting a mismatched one-off cell.
• Inspect every animation at normal pet size. Check direction, scale, silhouette, clipping, alpha holes, identity, and loop continuity. Numerical validation alone cannot prove that the pet looks right.

Finish and install

After the workflow's deterministic checks pass and the final previews have been visually reviewed, write qa/visual-review.json in the run:

{
  "approved": true,
  "standardAnimations": true,
  "lookDirections": true,
  "notes": "Describe what was actually checked."
}

Set these fields to true only after inspecting the corresponding output. Then run:

python "<skill>/scripts/pet_bridge.py" finish --run "<run>" --atlas "<run>/final/spritesheet.webp"

The bridge validates the version 2 atlas again, publishes pet.json and spritesheet.webp together, and marks the run ready. It preserves existing pets by choosing a new id if a name is already in use. Do not manually write an incomplete package into the live library.

The finished pet appears automatically in the Pets sidebar library. Tell the user its name and that Use pet activates it. Include the character/animation preview in the response. Selecting it changes the live pet; disabling the Pets plugin hides it and removes this skill's automatic registration while preserving created pets.