This server has a single tool with multiple critical quality gaps. The tool name 'run_gemini' is a generic verb that does not clearly convey action (get_, create_, search_, etc. are preferred). The description is in Korean and only 42 characters when translated ('Execute a prompt using Gemini'), which is below the rubric baseline of 194 chars average for tool descriptions. The tool combines multiple responsibilities: executing a prompt, analyzing files/directories/URLs, and managing working directories, this violates the single-responsibility principle. The input schema has three parameters with type definitions (all strings), but parameter descriptions are minimal and vague (e.g., 'Gemini에 전달할 프롬프트' provides no guidance on format, length, or constraints). No output schema is documented, the tool returns a dict with 'result' or 'error' keys, but field types and structure are not formally specified. Error handling returns generic error objects without actionable recovery guidance (e.g., 'FileNotFoundError' returns only 'Gemini CLI가 설치되어 있지 않습니다' without next steps). The tool takes a risk classification of WRITE but contains dangerous subprocess.run with shell=True and user-controlled prompt injection vectors (file_dir_url_path is concatenated directly into the shell command). No input validation, sanitization, or parameter constraints are present. The implementation is visible in the source, so this is not an inferred tool.

• Rename tool from 'run_gemini' to a clear action verb + noun pair. Examples: 'execute_gemini_prompt', 'analyze_with_gemini', or split into 'execute_prompt' and 'analyze_file'. Choose based on primary use case.
• Expand tool description to 100 - 200 characters explaining WHAT (executes Gemini with a prompt), WHEN (analyzing code, documents, etc.), and any prerequisites (Gemini CLI must be installed).
• Add detailed parameter descriptions: (1) 'prompt': specify expected format, max length, language, and that it will be augmented with file/directory context. (2) 'file_dir_url_path': clarify it accepts file paths, directory paths, or URLs; specify format and validation rules. (3) 'working_directory': explain it sets the execution context and must be a valid directory path.
• Document the output schema explicitly. Example: Returns object with fields: 'result' (string, stdout from Gemini), 'error' (string, if execution failed), 'duration_seconds' (number, execution time). Specify field types and when each field is present.
• Remove shell=True and use subprocess.run with shell=False, passing arguments as a list. This eliminates shell injection entirely. Example: subprocess.run(['gemini', '-m', 'gemini-2.5-flash', '-p', prompt], ...)
• Add input validation: (1) Validate working_directory exists and is accessible before changing to it. (2) Validate file_dir_url_path exists if it is a file/directory path; return actionable error if not. (3) Cap prompt length (e.g., max 4096 chars) with clear error message if exceeded.