The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input (text prompts, voiceover text, and image references) and interpolates them into commands executed by the genmedia CLI. It lacks robust sanitization to prevent malicious instructions embedded in user data from influencing the agent's behavior or tool calls.\n
Ingestion points: Untrusted data enters the context via user-supplied prompts for cinematography (cinematography.md), character identity (character-design.md), and speech text (character-lipsync.md, video-with-audio.md).\n
Boundary markers: Absent; the skill relies on simple prompt templates and does not use explicit delimiters or "ignore instructions" warnings for external content.\n
Capability inventory: The skill uses genmedia run for API/network operations and the --download flag for local file writing across multiple scripts.\n
Sanitization: No evidence of escaping or validation of user-provided strings before they are passed to tool arguments.\n- [COMMAND_EXECUTION]: The skill's primary purpose is the orchestration of multi-step pipelines using the genmedia CLI and shell utilities like jq to parse JSON outputs. This involves frequent execution of shell commands where user-controlled variables are interpolated into the command string.\n- [EXTERNAL_DOWNLOADS]: The skill downloads generated media from fal.ai endpoints to the local file system. These downloads are associated with the vendor's professional services and are documented as standard behavior for the skill's media production goals.

fal-recipes