The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple Ruby scripts (lib/buttercut/transcribe_job.rb, lib/buttercut/script_extractor.rb, lib/buttercut/process_footage.rb) using parameters derived from external inputs such as <video_path>, <transcript_output_dir>, and <language_code>. There is no evidence of input validation or shell escaping, which could allow for command injection if these variables contain shell metacharacters.
[DYNAMIC_EXECUTION]: The agent is instructed to perform complex file modifications using Edit and Grep tools on JSON transcript files. This logic involves runtime processing of external data to generate edit commands, increasing the risk of unintended file system operations if the input data is malicious.
[INDIRECT_PROMPT_INJECTION]: The refinement process in refine_instructions.md relies on analyzing and modifying transcripts generated from external audio files.
Ingestion points: Transcribed text from video files (via transcript_path), user_context, and footage_summary provided by the parent agent.
Boundary markers: The instructions include specific rules for 1-to-1 token replacement and word-count preservation, which act as limited constraints, but there are no explicit warnings to ignore instructions embedded in the audio transcript.
Capability inventory: The agent has the ability to read and write to the local filesystem and execute shell commands via the ruby scripts.
Sanitization: The skill lacks sanitization for the content of the transcript or context strings, meaning an attacker could craft audio that, when transcribed, attempts to manipulate the agent's behavior during the refinement step.

transcribe-audio