voice-scribe has 4 tools with complete input schemas and descriptions. Naming follows verb_noun convention (get_usage, transcribe, check_job, list_models). However, descriptions are brief (10-100 chars), parameter descriptions lack detail on constraints/formats, and output schemas are not documented. The transcribe tool has 14 parameters but many lack actionable guidance (e.g., 'prompt' has no format hint, 'speaker_threshold' lacks range). Error handling and recovery guidance are absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.

• Document output schemas for all tools. Example: transcribe returns {text: string, json?: object, job_id?: string, segments?: array}. Include field types and what each represents.
• Add min/max bounds and format hints to numeric/string parameters. E.g., 'speaker_threshold: float 0.0 - 1.0, default 0.5 (lower = more speakers detected)', 'offset_seconds: non-negative float, seconds from start'.
• Expand descriptions to 100 - 200 chars. Include WHEN to call (e.g., 'Call after transcribe returns a job_id to poll completion status'), WHAT it returns, and any prerequisites.
• Add tool annotations: transcribe with destructiveHint=true (initiates background job), check_job with idempotentHint=true (safe to retry), get_usage and list_models with readOnlyHint=true.
• Add error recovery guidance. E.g., 'If transcribe fails with "model not found", call list_models(scope="catalog") to see available models, then retry with a valid model name.'