The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to download a plugin from an unverified external source (github.com/Owl-Listener/designpowers.git) that is not identified as a trusted organization or well-known service.
[REMOTE_CODE_EXECUTION]: By following the installation steps to link or add the cloned repository as a plugin in tools like Claude Code or Gemini CLI, the agent is directed to execute external code in its runtime. The skill also relies on a custom designpowers CLI tool provided by this external source.
[COMMAND_EXECUTION]: The skill's documented workflow involves several shell commands for installation and operation, including git clone, claude plugin link, and the execution of the designpowers build and review commands.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, which provides a vulnerability surface for indirect prompt injection. 1. Ingestion points: The Review Lane accepts external URLs, screenshots, and code snippets for auditing. 2. Boundary markers: The instructions do not specify any delimiters or safety warnings to distinguish external data from agent commands. 3. Capability inventory: The skill possesses capabilities for shell execution, network access, and file system modification. 4. Sanitization: There are no documented procedures for sanitizing or validating the content retrieved from external sources during the audit process.

designpowers-agent-team