The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines specialized agent roles with specific communication constraints. Specifically, the Scribe and Ralph agents are instructed to work silently in the background or follow strict reporting patterns. In templates/scribe-charter.md, the instructions state, "Never speak to the user. Never appear in responses. Work silently." While these instructions match the design of a background logging service, they technically employ concealment patterns. This is assessed as a benign architectural choice for this framework.
[REMOTE_CODE_EXECUTION]: The skill implements a "runtime bootstrap" mechanism for its CLI tools. As described in decisions.md and the plugin-overlay-bundling skill, a PowerShell script (download.ps1) is used to resolve the latest release from GitHub assets and download self-contained Windows executables (excelcli.exe, mcp-excel.exe) for execution. This behavior is associated with the primary purpose of distributing the automation toolset and targets the author's verified GitHub releases.
[COMMAND_EXECUTION]: The coordinator agent defined in templates/squad.agent.md is granted capabilities to execute arbitrary shell commands to manage the development lifecycle. This includes using git for worktree management, gh for issue and pull request handling, and npm for dependency installation. These commands are executed within a local environment to perform repository maintenance tasks requested by the user.
[EXTERNAL_DOWNLOADS]: The framework frequently interacts with external resources, including cloning the author's distribution repository (sbroenne/mcp-server-excel-plugins) and fetching configuration samples for well-known services such as Trello, GitHub, and Notion as shown in mcp-config.md. These operations use official domains and established platform integrations.
[PROMPT_INJECTION]: An indirect prompt injection surface exists where the framework ingests data from GitHub issues and pull requests. As seen in templates/ralph-triage.js, external issue titles and bodies are processed to determine agent assignments. While this poses a surface for potential instruction injection, the system relies on structured role definitions and LLM guardrails to maintain operational integrity.

{skill-name}