The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's metadata contains instructions to install the summarize CLI using a third-party Homebrew tap (steipete/tap/summarize). This requires the agent to download and install binary software from a repository not managed by a recognized trusted organization.
[COMMAND_EXECUTION]: The skill instructions specify the execution of the summarize command with arguments provided by the user, such as URLs or local file paths. This relies on the agent's ability to run external binaries on the host system.
[PROMPT_INJECTION]: The skill is designed to process and summarize content from external, untrusted sources including web URLs, PDF files, and YouTube transcripts, creating a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through the output of the summarize command when it processes external URLs, local files, or YouTube content.
Boundary markers: No explicit delimiters or instructions are provided in the skill to warn the agent about potentially malicious instructions embedded within the content being summarized.
Capability inventory: The skill facilitates file reading and network access (via the summarize tool) and presents the resulting content to the LLM. No other dangerous capabilities were explicitly found in the skill's instructions.
Sanitization: There is no evidence of sanitization, filtering, or validation of the content retrieved from external sources before it is interpreted by the agent.

summarize