The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a wide range of CLI tools to interact with web services and social platforms. This includes standard binaries like curl and gh (GitHub CLI), specialized tools like yt-dlp (video content), and several platform-specific or custom utilities such as bili, twitter, rdt, opencli, and agent-reach itself.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and synthesize content from numerous external platforms (Twitter, Reddit, Bilibili, XiaoHongShu, etc.) and general web pages, which exposes the agent to untrusted data.
Ingestion points: Untrusted data enters the context through curl calls to services like Jina Reader (r.jina.ai) and V2EX, as well as the output of various social media search tools in SKILL.md.
Boundary markers: The skill provides no instructions or delimiters to help the agent distinguish between its core instructions and the potentially malicious data fetched from the internet.
Capability inventory: The agent has the ability to execute shell commands (curl, gh, yt-dlp, etc.) and write to the persistent directory ~/.agent-reach/.
Sanitization: The instructions do not define any sanitization, filtering, or validation processes for the data retrieved from external platforms.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch content from external sources, including the Jina Reader service and several public GitHub repositories for documentation and installation guides.
[REMOTE_CODE_EXECUTION]: The skill includes a procedure for checking and performing updates by fetching instructions from a remote repository. The agent is instructed to suggest a GitHub-hosted markdown file (update.md) to the user, which is then read and followed by the agent to perform the update. While requiring user initiation, this mechanism allows for the retrieval and execution of instructions defined in an external, remote location.

agent-reach