The Agent Skills Directory

[OBFUSCATION]: A hidden zero-width word joiner character (\u2060) was detected at the start of the Mapbox style URL string in rules/maps.md. This technique is often used to evade string-based security filters or hide content from automated scanners.
[EXTERNAL_DOWNLOADS]: The skill instructions depend on several external libraries, including an obscure package named mediabunny which is used for core media processing logic in multiple rule files but lacks explicit installation documentation.
[INDIRECT_PROMPT_INJECTION]: The skill provides code patterns for fetching and processing untrusted external data (such as subtitles, remote JSON metadata, and Lottie animations) without providing sanitization, validation, or boundary markers. Every indirect prompt injection finding requires documenting the evidence chain:
Ingestion points: rules/calculate-metadata.md (fetches from props.dataUrl), rules/import-srt-captions.md (fetches and parses .srt files), and rules/lottie.md (fetches Lottie JSON).
Boundary markers: None are present in the provided code examples.
Capability inventory: The skill uses network operations (fetch) and media parsing capabilities across its scripts.
Sanitization: None is implemented in the provided best practice examples.
[COMMAND_EXECUTION]: The skill provides numerous examples of shell commands for installing dependencies and adding framework features (e.g., npx remotion add, npm install), requiring the agent to execute code in the user's environment.

remotion-best-practices