The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions and templates for building extensions that ingest untrusted data, which creates a surface for indirect prompt injection attacks. \n
Ingestion points: Extensions process data through event handlers such as onAgentStarted, onPromptSubmitted, onToolCalled, and onRuleFilesRetrieved (defined in references/extension-interface.md). \n
Boundary markers: The provided templates in assets/templates/ do not include automated boundary markers or prompt delimiters for external content. \n
Capability inventory: Extensions have the capability to modify agent system prompts, inject context messages, block tool execution (e.g., onToolCalled), and interact with the project environment via the ExtensionContext (documented in references/event-types.md and references/extension-interface.md). \n
Sanitization: The skill does not enforce automated sanitization for external content, though it includes a reference to manual permission gating in references/examples-gallery.md. \n- [COMMAND_EXECUTION]: The skill provides documentation and examples for extensions that interact with system commands. \n
Evidence: The Permission Gate example in references/examples-gallery.md demonstrates how to monitor and intercept bash tool calls. \n- [EXTERNAL_DOWNLOADS]: The skill references external resources and dependencies as part of the extension development workflow. \n
Evidence: The skill fetches extension type definitions from the vendor's official GitHub repository (raw.githubusercontent.com/hotovo/aider-desk) and provides instructions for managing Node.js dependencies in folder-based extensions.

extension-creator