The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses instructional markers such as 'The rule' or 'Goal' to guide the agent's behavior, but these are used in a benign context to facilitate the writing partnership. No instructions were found that attempt to bypass safety filters or override system constraints in a malicious way.
[DATA_EXFILTRATION]: The skill manages local project files like outline.md, research.md, and style-profile.md to maintain state across sessions. There are no hardcoded credentials, sensitive file access (e.g., .ssh, .aws), or unauthorized network exfiltration patterns. It mentions using external tools for research, which is a standard and intended capability.
[REMOTE_CODE_EXECUTION]: The skill provides guidance on orchestrating with other agent skills (e.g., web-search-researcher, docs-seeker). These are references to standard agent capabilities and do not involve executing arbitrary remote code or shell commands from untrusted sources.
[COMMAND_EXECUTION]: No dangerous shell commands, privilege escalation (sudo), or persistence mechanisms were found. The skill's operations are limited to text processing and standard file I/O within the project scope.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided writing samples to learn the user's voice. While this is an ingestion point for untrusted data, the skill is designed to analyze stylistic markers (sentence rhythm, tone) rather than execute instructions within that data. The structured feedback loops and constant user check-ins act as inherent boundary markers.

writing-partner-pro