The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from live web searches, official documentation, and GitHub. It implements specific defensive instructions in rules/validation.md and rules/parallel-research.md that explicitly direct the agent to treat all retrieved content as evidence and ignore any instructions found within that content. This effectively mitigates the risk of external content hijacking the agent's behavior.
[DATA_EXFILTRATION]: The skill has the capability to fetch data from the web and read local repository files. However, its primary output is a markdown report saved to a local .hyper/research/ directory. It lacks any patterns associated with exfiltrating sensitive credentials or unauthorized network transmissions beyond the intended search and fetch functions.
[COMMAND_EXECUTION]: While the skill mentions preserving CLI commands in reports for documentation purposes, it does not include instructions or capabilities for arbitrary command execution. It explicitly restricts parallel subagents to a 'read-only research' mode, preventing unauthorized system changes.

research