The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to run the command python3 seedance_lint.py "<prompt text>" to validate user input. This pattern directly interpolates unvalidated user-supplied strings into a shell command, which allows for command injection if the user provides characters like ;, &, or | to execute arbitrary code.
[DYNAMIC_EXECUTION]: The instructions require the agent to execute external scripts, specifically seedance_lint.py and higgsfield_memory.py, which are expected to exist at the project root. Executing local scripts that process user-controlled input at runtime introduces a dynamic execution risk.
[PROMPT_INJECTION]: A central component of the skill is a 'Rewrite Playbook' and 'Voice Rewrite' guide specifically designed to bypass AI content filters. The instructions provide strategies to circumvent safety guardrails by substituting restricted terms (violence, brands, public figures) with descriptive archetypes that the filter might not detect.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data (user-generated scene descriptions) and pass it to execution environments. The instructions lack adequate boundary markers or sanitization logic for handling this external content.
Ingestion points: User-provided scene descriptions and failure reports processed in the 'Detailed Description' and 'Pre-flight Linter' sections.
Boundary markers: None identified. User input is directly interpolated into commands and prompt structures.
Capability inventory: Shell command execution via python3 for linting and memory logging (specified in SKILL.md).
Sanitization: No sanitization, escaping, or validation of user-supplied text is described before its use in shell execution or prompt construction.

higgsfield-seedance