The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The instructions in SKILL.md direct the agent to download a bash script (skills.sh) from the author's GitHub repository and execute it using the bash command to handle skill management logic.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill is designed to fetch SKILL.md instruction files from user-specified GitHub repositories, links, or local directories (see fetch and install functions in scripts/skills.sh).
Boundary markers: The script does not wrap the imported instructions in protective delimiters or provide warnings to the agent to ignore embedded instructions within the downloaded content.
Capability inventory: The skill uses curl, tar, ln, mv, and cp to manage system files and can execute commands on remote servers via ssh. The agent interprets the resulting downloaded files as authoritative instructions for future sessions.
Sanitization: No content validation or sanitization is performed on the downloaded SKILL.md files before they are installed and activated.
[COMMAND_EXECUTION]: The management script uses shell commands including curl, tar, ln, mv, and rm to manipulate files and directories in the user's home directory (~/.agents, ~/.claude). It also enables remote execution by piping the management script to a remote shell via ssh (e.g., ssh hermes 'bash -s -- install <source>' < skills.sh).

skill-manager