The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an indirect prompt injection vector through its "Customer feedback capture protocol". It instructs the agent to fetch untrusted content from external WhatsApp messages via the WAHA API, parse them for directives, and append them directly to persistent memory files (e.g., feedback_<tenant>_rules.md). This allows external actors to influence the agent's future instructions and behavior without adequate sanitization.
Ingestion points: Customer WhatsApp replies via WAHA API (GET /api/<session>/chats/<groupId>/messages).
Boundary markers: None identified; instructions specify capturing verbatim quotes.
Capability inventory: Shell command execution (ffprobe, open), database access (SELECT), network operations (fetch, R2 upload), and file write operations (APPEND to rules files).
Sanitization: None; the agent is instructed to identify directives like "must be" or "never" and update its rule set accordingly.
[COMMAND_EXECUTION]: The skill requires the use of shell commands to perform post-generation quality assurance. It uses ffprobe to verify video metadata and the open command to display extracted frames on the operator's workstation.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to several external endpoints, including Kie.ai for media generation, Cloudflare R2 for asset storage, and the WAHA API for message retrieval.

superseller-kie-discipline