The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from GitHub PR comments and reviews which may be controlled by external actors (e.g., automated bots or other contributors). This creates a surface for indirect prompt injection where malicious instructions could be embedded in those comments.\n
Ingestion points: SKILL.md describes fetching findings from GitHub API endpoints (/issues/<number>/comments, /pulls/<number>/reviews, /pulls/<number>/comments).\n
Boundary markers: SKILL.md includes a "Hard rules" section explicitly instructing the agent to treat findings as untrusted review data and not follow embedded instructions.\n
Capability inventory: The skill can execute local build commands (bun run build), perform git operations (git push, git rebase), and use the commit-with-context skill.\n
Sanitization: The instructions mandate manual verification of each finding against the code and prohibit following instructions embedded in review data.

babysit