The Agent Skills Directory

[COMMAND_EXECUTION]: Phase 8 of the pipeline instructs the agent to execute shell commands including cd, printf, and ffmpeg to concatenate video clips. These commands are executed directly in the local shell environment to finalize the production process.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data (the story text) which is used to generate a project_id. This identifier is then interpolated into shell commands and file paths, such as director_final_{project_id}.mp4 in Phase 8. The instructions do not include validation or sanitization for this variable, creating a vulnerability surface where a maliciously crafted story title could lead to command injection.
Ingestion points: User-provided story text and project metadata in Phase 1 (SKILL.md).
Boundary markers: None mentioned for the story or project_id variables.
Capability inventory: Shell command execution (ffmpeg, printf, cd) and file system write access for state files and concatenation lists.
Sanitization: Absent; the instructions do not specify filtering or escaping for user-derived variables before shell execution.
[DYNAMIC_EXECUTION]: The final assembly phase dynamically generates a configuration file (concat_list.txt) at runtime. This file is then used to control the execution of the ffmpeg tool, representing a pattern of generating executable or tool-directive content from runtime state.

director