ScreenMind has 7 tools with clear verb-noun naming (screenmind_watch, screenmind_record_start, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392). However, input schemas are present but lack depth: most parameters have type and description, but no enums for constrained values, no min/max bounds, and no output schemas documented. Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors. The tool set is well-composed (each does one thing), but descriptions could be more LLM-optimized (e.g., 'Process the latest screen recording' is vague about what 'process' entails). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (WRITE vs READ_ONLY marked in metadata but not in schema).

• Add output schemas to all tools. Document the structure of returned objects (e.g., screenmind_watch returns {summary: string, frames: [{timestamp: float, path: string, ocr_text: string}], transcript: string, timeline: [...]}). This enables LLMs to plan downstream calls and extract the right data.
• Add tool annotations: mark screenmind_record_start, screenmind_record_stop, and screenmind_config_set with destructiveHint=true; mark screenmind_watch, screenmind_list_recordings, screenmind_config_get with readOnlyHint=true. This helps agents reason about side effects.
• Add parameter constraints: declare 'since' as an enum or pattern (ISO_DATE | '7d' | '24h' | '30m'); add minValue/maxValue to 'max_frames' (e.g., 1-1000); add minValue/maxValue to 'dedup_threshold' and 'scene_change_threshold' (0-1).
• Enhance error handling: document what errors each tool can raise (e.g., 'ffmpeg not found', 'invalid video URL', 'OCR unavailable'), whether they are retryable, and what the LLM should do next. E.g., 'If OCR is unavailable, the tool returns frames without ocr_text, proceed with visual analysis.'
• Rewrite descriptions for LLM clarity: replace 'Process the latest screen recording' with 'Extract keyframes, OCR text, and audio transcript from the latest screen recording in capture_dir. Returns a text summary and file paths; Claude reads frames with the Read tool.' Include WHEN to use it (e.g., 'Call this after screenmind_record_stop to analyze a completed recording.').