The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides detailed reference guides and templates for configuring sub-agents with a high-privilege bypassPermissions mode. This mode is designed to skip all security prompts for tool execution, which represents a significant privilege escalation risk if applied to untrusted or externally-influenced agents.
[COMMAND_EXECUTION]: The reference guides document and provide examples for using lifecycle hooks (PreToolUse, PostToolUse) and MCP server definitions that allow for the execution of arbitrary shell commands. Examples include running validation scripts and using npx to execute packages.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates a high-risk workflow where it ingests natural language requirements to generate executable agent and skill definitions. This creates a surface for indirect prompt injection if the generated logic or system prompts are not properly sanitized.
Ingestion points: User-provided requirements for tool creation or modification processed in SKILL.md.
Boundary markers: The templates provided in the reference guides lack explicit boundary markers or instructions to ignore embedded commands in processed data.
Capability inventory: The generated tools are intended to have significant capabilities, including file system modification (Edit, Write) and shell access (Bash).
Sanitization: There are no instructions for the AI to sanitize user-provided descriptions or logic before incorporating them into the generated tool's system prompts.
[PERSISTENCE]: The skill enables and documents the configuration of persistent memory scopes (user, project, local). This allows agents to maintain state, learned patterns, and architectural knowledge across different sessions and projects.

edit-tool