The Agent Skills Directory

[PROMPT_INJECTION]: The file references/bash-agent-notes.markdown contains instructions designed to manipulate the agent's security reasoning. It explicitly tells the agent to 'avoid knee-jerk reactions,' stop the 'cargo cult' of flagging eval as dangerous, and adopt a 'trusted actor' model for specific internal projects. These instructions attempt to override standard security auditing best practices by substituting them with a more lenient, context-dependent posture that is unverifiable at runtime.
[COMMAND_EXECUTION]: Multiple files, including references/code-examples.md and references/bash-agent-notes.markdown, provide patterns for executing arbitrary shell commands via eval and sudo. While these are framed as functional examples, the accompanying instructions to minimize scrutiny of these patterns increase the risk of the agent generating or approving dangerous code.
[EXTERNAL_DOWNLOADS]: The skill references an external repository (github.com/dylanaraps/pure-bash-bible) for extensive bash patterns. While the source is a well-known community resource, the skill encourages reliance on external code templates for core logic.

bash-development