The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads binaries and AI model weights (such as Roto Brush and Face Tracking models) from external sources, including https://github.com/storytold/effectcraft. Although it employs SHA-256 integrity verification, the sources are not verified trusted organizations.
[REMOTE_CODE_EXECUTION]: Through bootstrap.py and workflow.py, the skill automates the retrieval and execution of the EffectCraft CLI and its dependencies.
[DYNAMIC_EXECUTION]: Multiple entry points, such as script.run, run_script, and file.executeFile, allow for the execution of arbitrary JavaScript code or project-embedded scripts within the application context. This presents a path for executing potentially malicious logic if provided with untrusted project data.
[COMMAND_EXECUTION]: Management scripts including launch.sh (macOS/Linux) and launch.ps1 (Windows) are used to execute shell commands for environment setup and tool orchestration.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a vulnerability surface for indirect prompt injection due to its handling of external data.
Ingestion points: It processes structured data from JSON workflow plans (e.g., brand-intro.json) and binary project files (.ecproj).
Boundary markers: The instructions lack specific delimiters or guidance to ignore natural language instructions that might be embedded within these data formats.
Capability inventory: The environment includes high-privilege capabilities such as arbitrary script execution, file system operations, and network access for downloads.
Sanitization: The skill relies on schema structure validation, which does not effectively filter for malicious natural language directives hidden in data fields.

effectcraft-cli-setup