Three tools with clear, action-oriented names (video_listen, video_see, watch_video) and well-written descriptions (150-180 chars each) that explain WHAT, WHEN, and WHY. Input schemas are present with proper types and descriptions. However, output schemas are completely undocumented, LLMs cannot plan downstream operations or extract fields reliably. Error handling is minimal (generic success/error flags). No parameter validation guidance, no recovery hints, and no mention of rate limits or timeout behavior. Descriptions are good but lack dependency hints (e.g., 'Call video_listen first if you only need audio'). Overall solid naming and description work, but missing critical schema documentation and error guidance.

• Document output schemas explicitly. For each tool, specify: {success: boolean, duration_seconds: number, frame_count?: number, frames?: [string (base64 JPEG)], transcript?: string, url: string, error?: string}. Include field types and when each field is present.
• Add error classification to responses. Return {success: false, error: string, error_type: 'retryable'|'user_fixable'|'fatal', recovery_hint: string}. E.g., 'error_type: retryable, recovery_hint: Network timeout. Retry in 30 seconds.'
• Constrain max_frames parameter. Update description to: 'Maximum frames to extract (1-20, default 5). Higher values increase processing time and token usage.'
• Add timeout and rate-limit guidance to tool descriptions. E.g., 'Timeout: 300 seconds. If video is longer than 2 hours, consider calling video_listen first to check duration.'
• Include cost/performance hints in descriptions. E.g., video_listen: 'Lightweight, no GPU required, ~10 seconds for typical videos.' watch_video: 'Full analysis, requires GPU, ~30-60 seconds for typical videos.'