The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from video transcripts and user-provided posts (Step 0 in SKILL.md) to generate articles and infer writing styles. A malicious transcript could contain instructions designed to influence the agent's behavior during the writing or verification process.
Ingestion points: Transcript files (e.g., rag.json, transcript.txt) and user posts for voice profile creation.
Boundary markers: The skill uses ---- and ==== separators for output structure, but does not provide explicit boundary instructions for the agent when reading input transcripts.
Capability inventory: The skill executes shell scripts (fetch_logo.sh, render_cover.sh), writes to local configuration files (~/.claude/voice-profile.md), and runs a local HTTP server.
Sanitization: No sanitization is performed on transcript text before processing.
[EXTERNAL_DOWNLOADS]: The scripts/fetch_logo.sh utility uses curl to fetch logo files from external domains provided by the user. While it includes timeouts and size checks, it allows the agent to make network requests to arbitrary external websites.
[COMMAND_EXECUTION]: The skill relies on executing several shell scripts to perform its tasks. render_cover.sh manages a local Python HTTP server and searches the user's home directory for Playwright screenshots, while fetch_logo.sh executes complex curl and grep pipelines based on user input.
[DATA_EXPOSURE]: The skill is designed to read from and write to ~/.claude/voice-profile.md. While this is intended for personalizing content, it involves the agent accessing and modifying files within the application's configuration directory.

video-to-articles