The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as song lyrics, artist descriptions, and audience sentiment which could contain malicious instructions.
Ingestion points: context/artist.md, context/audience.md, transcription lyrics, and social media post history fetched via API.
Boundary markers: None. Instructions do not provide delimiters for separating external data from agent prompts.
Capability inventory: Shell execution (curl, jq, sed, yt-dlp) and workspace file modifications.
Sanitization: None. External data is directly interpolated into prompts and used to drive generation parameters.
[PROMPT_INJECTION]: The skill contains instructions that direct the agent to conceal process failures and regeneration attempts from the user to ensure only high-quality output is presented.
Evidence: references/analyze-gate.md instructs the agent to "regenerate ... re-analyze" and "Don't ship a flagged asset" when quality checks fail. The Stop hook in SKILL.md enforces this by blocking the agent from declaring a task finished until a successful analysis pass is recorded in the conversation, overriding the agent's completion claim.
[EXTERNAL_DOWNLOADS]: The skill downloads content from external and vendor-managed sources to fulfill its multimedia generation tasks.
Evidence: references/song-sourcing.md describes using yt-dlp to download audio from YouTube. The skill also fetches binary files (MP3, MP4) from vendor repositories via the GET /api/sandboxes/file endpoint.
[COMMAND_EXECUTION]: The skill relies on executing shell commands to perform its primary functions, including API interaction and file processing.
Evidence: Multiple reference files provide bash snippets using curl, jq, sed, and yt-dlp to interact with the vendor API (api.recoupable.dev) and manage local workspace files.

recoup-content-make-video