Scoring was not performed

• Document output schemas for all 17 tools. For transcribe_audio/transcribe_video, specify that response includes {text: string, segments: Array<{start_ms, end_ms, text}>}. For info tools, document returned fields like {duration, format, codec, bitrate}.
• Add explicit enum constraints to parameters: model_size ∈ [tiny, base, small, medium, large], lang ∈ [en, th, ja, zh, ko, ...], style ∈ [bullet_points, paragraph, key_takeaways, detailed], tone ∈ [formal, casual, technical], format ∈ [srt, vtt, ass, txt, json].
• Clarify mutual exclusivity in dual-input tools: 'Provide either file_path (local file) OR url (remote file), not both. If both are provided, file_path takes precedence.'
• Add error recovery guidance: 'If transcription fails with a timeout, try smaller model_size (tiny or base). If youtube_to_text fails, verify the video ID is valid (12-character alphanumeric).'
• Document idempotency contracts: 'Calling batch_transcribe twice with identical sources produces identical output files. Existing files are overwritten without warning.' Or provide a skip_existing parameter.
• Add tool-chaining guidance in summarize_transcript: 'Call transcribe_audio or transcribe_video first to obtain transcript text. Then pass result.text to this tool.' Similar hints for generate_blog_post, extract_key_moments.
• Standardize abbreviations: rename vdo_info to video_info for consistency with transcribe_video.
• Add pagination/limit guidance: If batch_transcribe or transcribe_folder can return hundreds of results, document max result count and provide offset/limit parameters.