The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the muapi CLI tool (e.g., muapi image edit, muapi predict wait) and curl for interacting with the MuAPI service.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to api.muapi.ai via curl to submit and poll for media generation tasks when CLI aliases are not available.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by interpolating untrusted user-supplied data (product_name, human_image, product_image) directly into prompts for image, script, and video generation models.
Ingestion points: Inputs defined in SKILL.md (product_name, human_image, product_image) are used in Phase A and Phase B prompts.
Boundary markers: No explicit delimiters or instructions are used to prevent the agent from obeying instructions embedded within the user-provided product details or image metadata.
Capability inventory: The skill has the capability to execute shell commands (muapi, curl) and perform network requests, which increases the impact if an injection occurs.
Sanitization: No sanitization or validation of the input strings or URLs is performed before they are interpolated into the system prompts.
[CREDENTIALS_UNSAFE]: The skill instructions direct the agent to handle a sensitive credential (MUAPI_API_KEY) by passing it through shell environment variables and HTTP headers (-H "x-api-key: $MUAPI_API_KEY"). While using environment variables is a common practice, the agent is actively managed to use this secret in shell contexts.

muapi-ugc-ads-workflow