All three tools have minimal descriptions (under 50 chars) and identical single-parameter schemas with no type constraints. Descriptions lack WHAT/WHEN/WHY context required for LLM tool selection. No output schemas documented. Tool names lack action verbs (extract_document, classify_document, summarize_document are acceptable but descriptions are too sparse). Parameters have type but no constraints, ranges, or format guidance. No error handling guidance. Tools return JSON strings instead of structured objects, forcing LLM parsing. Billing integration is hardcoded, not abstracted. No pagination, no idempotency hints, no recovery guidance.

• Expand each tool description to 100-200 characters. Example: 'Extract text from a PDF or image using OCR. Returns extracted text, confidence scores, and detected language. Use this to read document content before classification or summarization. Supports PDF, PNG, JPEG up to 50MB.'
• Document output schema for each tool. Example: extract_document returns {"text": string, "confidence": number 0-1, "language": string, "pages": number, "file_type": string}. Classify_document returns {"type": enum[invoice|report|contract|...], "confidence": number, "reasoning": string}.
• Add parameter constraints: 'path (string, required): Absolute or relative file path. Supported: .pdf, .png, .jpg, .jpeg. Max 50MB. Example: /documents/invoice_2024.pdf'
• Replace JSON string responses with structured objects. Return {"success": true, "data": {...}, "credits_remaining": number} instead of JSON-encoded strings. This eliminates LLM parsing overhead.
• Add error classification and recovery guidance. Example: 'File not found at /path/to/file. Verify the path exists and is readable. Supported formats: PDF, PNG, JPEG. Max size: 50MB.' For billing errors: 'Out of credits. Remaining: 0. Upgrade at https://... or retry tomorrow (daily limit resets at 00:00 UTC).'