The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes an instruction to the agent that any version of the 'redline' skill found within the local repository directory (.claude/skills/redline/ or .agents/skills/redline/) should be considered authoritative and followed instead of the global version. This creates a mechanism for untrusted local files to override the agent's core instructions and safety parameters.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from a Markdown file located in the user's downloads folder and applies the changes directly to the source code or documents.
Ingestion points: Reads the latest *.redline.md file from ~/Downloads (Linux/macOS) or %USERPROFILE%\Downloads (Windows).
Boundary markers: None identified; instructions emphasize that human edits are 'decisions' to be applied 'verbatim'.
Capability inventory: The agent has the capability to write to the local file system to apply the reviewed changes.
Sanitization: No sanitization or validation of the input file's content is performed before application to the source artifact.
[COMMAND_EXECUTION]: The skill triggers platform-specific shell commands to open generated HTML files in the user's default browser (open, xdg-open, start).
[COMMAND_EXECUTION]: The skill performs shell-based file system lookups to identify the newest review file in the downloads directory using ls or dir commands.

redline