The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external inputs such as error messages, system logs, API responses, and database contents across several steps (e.g., Phase 1.1, 1.4, 5c, and 6b). Because the agent is also instructed to execute shell commands based on its analysis, it is vulnerable to indirect prompt injection if malicious instructions are embedded in the data being debugged.
[COMMAND_EXECUTION]: The skill utilizes several shell commands for diagnostics and environment management, including pytest, git, sqlite3, and pip. It also includes macOS-specific commands like tccutil and rm -rf for resetting application state, which should be used with caution as they modify system privacy settings and delete user data in application containers.
[REMOTE_CODE_EXECUTION]: The skill reference guides suggest using pip install to ensure the correct version of a dependency is installed from PyPI. This is used to resolve issues with incorrect local forks of libraries and follows standard development practices.

systematic-debugging