The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to execute powerful system commands to identify, archive, and delete directories across the user's filesystem.
Evidence: Employs find, rm -rf, and tar on POSIX systems, and Get-ChildItem, Remove-Item -Recurse, and [System.IO.Directory]::Delete on Windows.
Evidence: Performs recursive deletion of directory trees and removal of symbolic links or junctions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the local filesystem to influence its internal logic and decision-making.
Ingestion points: In Phase 2, the agent is instructed to read the SKILL.md frontmatter from arbitrary directories found during the search phase to resolve the 'keep list'.
Boundary markers: The instructions lack specific sanitization steps or delimiters to isolate the metadata being read from external skill files.
Capability inventory: The agent possesses the capability to execute shell commands, write to the filesystem (backups), and perform recursive deletions.
Sanitization: No validation or escaping is described for the metadata.documentation field or folder names before they are used to determine which items are kept or deleted.
[DATA_EXFILTRATION]: The skill performs an automated sweep of the user's home directory to discover environment details.
Evidence: Searches the user's $HOME or %USERPROFILE% directory up to a depth of 3 levels to locate all directories named skills.
Evidence: Uses git rev-parse --show-toplevel on discovered paths to map repository structures and work trees.

awesome-skills-purge