The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from external YouTube sources, creating a potential surface for indirect prompt injection.
Ingestion points: hits.chunks containing transcript text in the JavaScript block in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded content are used in the prompt templates.
Capability inventory: The skill utilizes arcmira_execute_read and arcmira_execute_write for executing logic and potentially modifying state.
Sanitization: Transcript text is truncated to 300 characters (c.text.slice(0, 300)), but no structural sanitization or escaping is applied to prevent the model from obeying embedded commands.
[DYNAMIC_EXECUTION]: The skill relies on executing JavaScript code snippets at runtime via the arcmira_execute_read tool.
Evidence: The "Worked program" section in SKILL.md provides code to be executed for searching and processing results.
Context: This execution is part of the standard operating procedure for the Arcmira platform and follows predefined templates.

podcast-transcript-search