The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of transcripts and raw recording metadata to generate structure plans and edit decision lists. This represents a potential surface where malicious content embedded in transcripts could attempt to influence the agent's logic.
Ingestion points: Recording transcripts and metadata ingestion in SKILL.md.
Boundary markers: None explicitly defined in the example prompts to distinguish transcript content from instructions.
Capability inventory: The skill generates and executes FFmpeg shell commands, Bash scripts, and Remotion (React/Node.js) code.
Sanitization: No explicit sanitization or validation of transcript content before interpolation into prompts.
[DYNAMIC_EXECUTION]: The skill workflow involves 'scaffolding' code, where an LLM generates Remotion components and FFmpeg command sequences which are subsequently executed (npx remotion render, bash cuts.sh). This dynamic generation and execution of scripts is inherent to the skill's primary purpose but is documented as a known capability.
[COMMAND_EXECUTION]: The skill relies extensively on local shell command execution for video processing, including complex FFmpeg pipelines for cutting, merging, and reframing footage (SKILL.md).

video-editing