The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies an attack surface by ingesting external brand documentation files (brand-identity.md, ICP.md, and messaging.md) to generate video storyboards and director prompts.
Ingestion points: Files are read from the working directory or a brand/ subdirectory as specified in SKILL.md (Step 1).
Boundary markers: The instructions lack explicit delimiters for these files, however, the skill focuses on extracting specific thematic data.
Capability inventory: The skill can execute shell commands (bash), write files (downloaded images), and perform network operations (curl) via the run-social-video.sh script.
Sanitization: The orchestrator script uses shell array expansion ("${SEEDANCE_ARGS[@]}") when invoking sub-processes, which effectively prevents simple argument splitting or command injection from the generated prompt content.
[COMMAND_EXECUTION]: The skill uses a bash script run-social-video.sh to coordinate calls to other internal tools like generate-image.sh and generate-seedance.sh. The script uses set -euo pipefail for error handling and correctly handles variable quoting to maintain execution integrity.
[EXTERNAL_DOWNLOADS]: The script uses curl to download reference images generated by the image model. The target URLs are dynamically retrieved from the output of the internal generate-image.sh tool rather than being hardcoded or retrieved from untrusted user input.
[CREDENTIALS_UNSAFE]: The skill manages its API authentication by sourcing a .env file from the repository root to load MUAPI_KEY. This follows industry best practices for secret management and does not involve hardcoded credentials.

muapi-social-media-video