The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute a local linter script (seedance_lint.py) and interpolate generated prompt content directly into the shell command. This pattern is vulnerable to command injection if the prompt content includes shell metacharacters or maliciously crafted strings.
Evidence: Located in the 'Worked Examples' section: python3 seedance_lint.py --preflight --model seedance_2_0 "<prompt>". While quotes are used, they do not provide sufficient protection against complex injection payloads.
[PROMPT_INJECTION]: The skill uses template-based placeholders that ingest untrusted user data, creating a surface for indirect prompt injection.
Ingestion points: Placeholders such as [CHARACTER] and [PASTE ONLY THE AUs] in SKILL.md which are intended to be filled with user-supplied content.
Boundary markers: The skill uses simple bracketed delimiters but lacks explicit instructions or warnings to the model to ignore or sanitize embedded commands within these blocks.
Capability inventory: The skill explicitly invokes shell command execution capabilities (python3).
Sanitization: There are no documented instructions for escaping or validating input before it is used in prompt construction or linter execution.
[EXTERNAL_DOWNLOADS]: The skill references an external scientific resource for verification of Action Unit codes.
URL: melindaozel.com/facs-cheat-sheet.
Context: Provided as a reference for users and the agent to cross-check the accuracy of generated FACS Action Units.

higgsfield-facs