The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input (character descriptions and scene prompts) which is processed by the agent to generate ComfyUI workflows and execute scripts.
Ingestion points: User prompts for character generation and test case data in eval/test-cases.yaml.
Boundary markers: Absent. The instructions do not define delimiters or specific warnings to ignore instructions embedded within user-provided content.
Capability inventory: The skill includes Python scripts (references/talking-head-workflows.md) using subprocess.run and an evaluation script (eval/run-eval.sh) that executes the claude CLI.
Sanitization: Absent. User input is used to form prompts and command-line arguments without visible validation or escaping.
[COMMAND_EXECUTION]: Scripts provided within the skill execute external shell commands and system tools.
eval/run-eval.sh executes the claude CLI along with standard utilities like grep, wc, python3, and node to manage evaluation tasks.
references/talking-head-workflows.md contains Python scripts that use subprocess.run to call external inference scripts for audio and video processing, such as inference.py and inference_codeformer.py.
[EXTERNAL_DOWNLOADS]: The skill provides comprehensive instructions for downloading third-party code and large-scale models from external sources.
References numerous GitHub repositories, HuggingFace repositories, and Civitai models.
Provides explicit commands for git clone and pip install targeting external repositories including SadTalker, LivePortrait, and Wav2Lip.
References established research organizations and technology companies in the AI ecosystem.

comfyui-character-gen