The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/clip-finder.mjs executes shell commands using node:child_process.execSync. The commands are parsed from clip-plan.md by identifying lines that start with ffmpeg. Because the content of the plan file is derived from external transcript data and can be influenced or modified by the agent, there is a risk of command injection if the file is manipulated to include additional shell operators or malicious commands.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted podcast and video transcripts. These transcripts are parsed and then presented to a subagent for auditing and creative judgment. A malicious transcript could contain hidden instructions or adversarial content designed to override the agent's logic or bypass the safety and quality rails defined in the skill instructions.
Ingestion points: Untrusted data enters the environment through transcript.txt and is processed by scripts/clip-finder.mjs and the main agent.
Boundary markers: No boundary markers or 'ignore' instructions are used when passing transcript content to the subagent auditor in templates/clips-auditor-brief.md.
Capability inventory: The skill has access to shell execution via execSync in scripts/clip-finder.mjs and file system modification through node:fs methods across its scripts.
Sanitization: The scripts do not perform any sanitization, escaping, or validation on the transcript text before it is interpolated into files or prompts.

podcast-to-shorts