The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install a package named findskill and additional skill modules from an external, non-whitelisted registry.
[REMOTE_CODE_EXECUTION]: The instructions encourage the use of npm install -g and npx to fetch and execute code from a remote repository, providing a path for arbitrary code execution if the source package is compromised.
[COMMAND_EXECUTION]: The agent is directed to execute various shell commands to manage the skill lifecycle, including installation, updates, and reading external file content.
[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to read and follow instructions contained within downloaded SKILL.md files, creating a significant attack surface for indirect prompt injection.
Ingestion points: cat ~/.claude/skills/<skill-name>/SKILL.md in the agent workflow.
Boundary markers: None present; the agent is explicitly told to follow the instructions in the downloaded files without isolation.
Capability inventory: The agent maintains shell access, npm package management, and file system write capabilities while processing untrusted instructions.
Sanitization: No sanitization or verification of the external markdown content is performed before the agent interprets it as instructions.

use-findskill