The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads from sensitive configuration paths including ~/.claude.json and .mcp.json to extract skill and server metadata, which may contain authentication details.
[COMMAND_EXECUTION]: The skill provides shell commands for the user to execute that perform recursive directory deletion (rm -rf) for user-scoped skills and modify core configuration files.
[CREDENTIALS_UNSAFE]: By accessing and managing ~/.claude.json, the skill interacts with a file that often stores environment-specific credentials and API tokens.
[COMMAND_EXECUTION]: The skill instructions prompt the agent to deliver commands meant to be run with the ! interactive terminal prefix, which bypasses the agent's restricted tool environment.
[INDIRECT_PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes untrusted data from the SKILL.md files of all installed skills.
Ingestion points: Reads SKILL.md and directory metadata of every installed skill.
Boundary markers: No explicit boundary markers or ignore-instructions are specified for the TUI display.
Capability inventory: Filesystem deletion (rm -rf), configuration file modification (~/.claude.json).
Sanitization: No sanitization of the skill descriptions or metadata is described before display in the TUI.

janitor-swipe