The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to research and process content from arbitrary external sources, including URLs, GitHub repositories, and npm packages. This ingestion of untrusted data creates a surface for indirect prompt injection, where malicious instructions embedded in a researched document could influence the agent's behavior. While the skill includes extensive methodologies for verification (Chain-of-Verification, Atomic Fact Decomposition) to prevent hallucinations, it does not mandate technical boundary markers (e.g., specific delimiters) for the ingested content. The agent's access to shell tools (gh, git) increases the potential impact of such an injection.
Ingestion points: WebFetch, tavily_extract_process, ai_search, github_search, and git clone operations across multiple files.
Boundary markers: The instructions emphasize quoting and citing but do not provide programmatic delimiters to separate untrusted content from system instructions.
Capability inventory: Access to bash shell, GitHub CLI (gh), and Git.
Sanitization: The skill focuses on factual verification and atomic fact decomposition to ensure accuracy, but lacks specific sanitization for adversarial injection patterns in the data.
[EXTERNAL_DOWNLOADS]: The skill fetches data from several external services, including an unofficial and undocumented API at npmx.dev for npm package documentation. It also references unpkg.com and the npm registry for fallback research.
[COMMAND_EXECUTION]: The skill provides patterns for executing shell commands, specifically gh api for fetching repository contents and git clone for local source inspection. These tools are used to interact with data fetched from external repositories.

research