The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data, such as text prompts and images, which are then passed to 3D generation tools and a Python-based scene builder.
Ingestion points: User-provided text descriptions (e.g., for meshy_v6_text_to_3d) and image references (e.g., for image_to_3d) handled via the instructions in SKILL.md.
Boundary markers: The instructions provide specific parameter mapping rules and emphasize that geometry changes should be handled in the source image stage ("fix it in 2D, then lift it").
Capability inventory: Access to 3D generation tools (generate_3d) and a Blender scene builder tool that executes Python code (scene_builder_3d_run_python).
Sanitization: The skill contains explicit safety constraints for the scene builder, prohibiting network fetches and embedded model bytes within the Python environment.
[COMMAND_EXECUTION]: The skill defines workflows for the scene_builder_3d_run_python tool, which executes Python commands within a Blender 5.2 environment to build 3D scenes.
Evidence: The "3D Jutsu" section details the execution of Python against a private Blender project, providing guards like revisionBefore to ensure state consistency during mutations.
[EXTERNAL_DOWNLOADS]: The skill references local specifications and scripts within the repository structure for data synchronization and model verification.
Evidence: References to ../../specs/models_explore_snapshot_3d_2026-09-26.json and ../../scripts/sync_specs.py. These are internal file system references within the vendor's repository structure and do not trigger external network downloads.

higgsfield-3d