The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted codebase content during its audit phase. It successfully mitigates this risk by defining 'Hard Rule 4', which instructs the agent to treat repository content as inert data and to ignore any embedded instructions (e.g., 'ignore previous instructions').
[PROMPT_INJECTION]: The skill includes text that mentions common injection patterns as examples for the agent to watch for and report. These are identified as defensive instructions rather than malicious attempts to override the agent's behavior.
[COMMAND_EXECUTION]: While the skill performs reconnaissance using tools like grep, it explicitly forbids mutating operations such as package installations, builds, or commits. The 'execute' variant is scoped to operate in an isolated worktree, providing a layer of protection for the primary codebase.
[DATA_EXPOSURE]: The skill is designed for source code analysis and does not access sensitive user credentials, environment variables, or perform network exfiltration.

improve-animations