Three tools with complete input schemas and descriptions, but descriptions are generic and lack LLM-optimized guidance. All tools have proper Zod validation with constraints (min/max length, numeric bounds). However, descriptions do not explain WHEN to use each tool, dependencies, or recovery paths. Output schemas are partially documented in code but not formally declared. No error handling guidance for downstream service failures. Tool names follow verb_noun convention correctly.

• Expand tool descriptions to 150-250 chars. Add WHEN to use: 'Use to convert text to natural-sounding speech via ElevenLabs. Supports multiple voices and models. Returns audio URL.' Include prerequisites: 'Requires valid ElevenLabs voiceId and API token configured server-side.'
• Document output schemas in tool metadata or comments. E.g., 'Returns {content: [{type: "audio"|"image"|"text", url?: string, text?: string, mime_type?: string}]}' so LLMs know what fields to extract.
• Add error recovery guidance. E.g., 'If service returns 429 (rate limit), wait 60s and retry. If 401 (auth), check TTS_TOKEN env var. If 500, service is down, try again in 5 minutes.'
• Enumerate valid voiceId and modelId values in descriptions or as enums. E.g., 'modelId must be one of: eleven_turbo_v2, eleven_multilingual_v2' to prevent hallucinated values.
• Document non-idempotency for generative tools. Add note: 'Each call produces different output. Do not retry on transient errors without user approval, as it will generate new content.'