Single tool 'readtext' has a clear verb-noun name and acceptable description (89 chars), but critical gaps undermine quality. Input schema is visible with two parameters (image_path, detail), both typed as string/boolean with descriptions. However, the output schema is completely undocumented, the description mentions 'text lines with bounding boxes and confidences' but no structured output format is defined. This forces LLMs to guess at response structure. The 'detail' parameter defaults to true but lacks explanation of what 'full detail' vs summary mode returns. No error handling guidance (e.g., what happens if image_path is invalid, unsupported format, or OCR fails). No pagination or result limits documented despite OCR potentially returning many lines. Missing: idempotency statement, permission/scope declaration, and recovery guidance for common failures.

• Document the complete output schema. Example: 'Returns {"lines": [{"text": string, "confidence": number 0-1, "bbox": {"x": number, "y": number, "width": number, "height": number}}], "total_lines": number}' when detail=true, or '{"lines": [string]}' when detail=false.
• Add error handling guidance to description: 'If image_path is invalid or unsupported format, returns error with suggestion to check file exists and is PNG/JPEG/BMP/GIF/TIFF/WebP. If OCR fails, returns error with confidence threshold info.'
• Clarify the 'detail' parameter: 'When true, returns text, confidence score (0-1), and bounding box per line. When false, returns text lines only. Defaults to true.'
• Document result limits: 'Returns up to 1000 text lines per image. For images with more lines, consider splitting the image or using a higher confidence threshold.'
• Add idempotency statement: 'This tool is idempotent, repeated calls with the same image_path return identical results.'