The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides explicit behavioral instructions to steer the agent's persona and response style. Modules like 'Act, don't overplan' and 'Stop only at real boundaries' instruct the agent to prioritize speed and autonomy by bypassing confirmation for non-destructive tasks. However, these instructions are balanced by 'Guard rules' that explicitly forbid hallucinating results and require tool verification.
[INDIRECT_PROMPT_INJECTION]: By modifying how the agent responds to user input (e.g., 'Lead with the outcome', 'Match effort to the task'), the skill creates a surface where the agent's decision-making process is influenced by the instructions when processing external data. The document provides clear mitigations for this, such as the 'Judgment Wall' and 'Agency Wall' sections which advise the agent to challenge incorrect premises and admit capability limits.
[SAFE]: The skill is distributed as a single Markdown file containing only text-based instructions. It does not include any executable scripts, shell commands, network requests, or external package dependencies.

operator-brain