The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted user data via <user_prompt>, <examples>, and <feedback> to generate new instructions. It does not include explicit instructions to sanitize these inputs or ignore embedded commands during its internal analysis steps. Specifically, there are no boundary markers defined to prevent the model from executing instructions contained within the user-provided prompt data. The skill possesses the capability to write and delete files within the repository's docs/ workspace, which could be abused if malicious instructions are followed.\n- [COMMAND_EXECUTION]: The skill performs file system operations using computed paths that are influenced by user input. During Phase D, the skill deletes a draft file at a path incorporating the <prompt-name>, which is determined based on the content of the user's initial prompt. Without strict validation or sanitization of this generated name by the agent, there is a potential risk of path traversal or unintended file deletion if a malicious input influences the filename construction.\n- [EXTERNAL_DOWNLOADS]: The skill configuration in agents/openai.yaml includes a dependency on the openaiDeveloperDocs MCP server located at https://developers.openai.com/mcp. This reference targets a well-known service from a trusted organization and is utilized for retrieving technical documentation.

ralph-prompt-generator