The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repository files and live webpages, which are defined as ingestion points in the 'Trust Boundary' section of SKILL.md. This section provides explicit boundary markers by instructing the agent to ignore any embedded instructions that attempt to expand scope or redirect publication. The skill has capabilities for command execution and network operations, but mandates sanitization by excluding authentication tokens and private data from all captures.\n- [DYNAMIC_EXECUTION]: The skill is designed to discover and execute repository-native capture commands and utilizes browser-testing frameworks. Instructions in SKILL.md direct the agent to read repository documentation to find existing capture contracts and to use the repository's Playwright stack for automation.\n- [DATA_EXFILTRATION]: The skill has functionality for sharing media to Slack and publishing to Figma. This risk is mitigated by specific policies requiring the agent to obtain explicit user identification of the destination and user confirmation before performing any external write or upload.\n- [COMMAND_EXECUTION]: The skill utilizes external command-line tools such as 'ffprobe' to verify media metadata and pixel formats during the bundle generation process.

capture-feature