The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by ingesting and processing untrusted content from the internet into the agent's context.
Ingestion points: Step 1 instructions in SKILL.md direct the agent to use WebSearch to gather information from external sites like X/Twitter, Reddit, and various technical forums.
Boundary markers: There are no boundary markers or instructions to treat the retrieved content as data rather than instructions, which could allow malicious commands embedded in search results to override the agent's behavior.
Capability inventory: The skill is designed for use in environments like Claude Code, which have the capability to read/write files and execute commands.
Sanitization: The workflow lacks any sanitization or validation steps to filter potential malicious payloads from the fetched web content before it is summarized or used for writing.
[EXTERNAL_DOWNLOADS]: The skill relies on the WebSearch tool to retrieve data from external sources. While this is intrinsic to its function, the reliance on unverified third-party platforms for content generation introduces a risk of processing malicious data.

wechat-article-writer