Two tools with minimal descriptions (both under 100 chars) and schemas present but sparse. insert_prefix and lookup_prompt lack parameter descriptions entirely, the schema shows property names but no 'description' fields for 'prompt' or 'kv_cache_id'. No output schema documented. No error handling guidance. Tool names are verb-noun but descriptions are vague about when/why to use them. No mention of side effects, prerequisites, or recovery paths. Schemas are valid JSON Schema but lack the richness needed for LLM reasoning.

• Add parameter descriptions to both tools. For 'prompt': 'The text to cache or lookup. Whitespace-separated tokens. Max 10,000 chars.' For 'kv_cache_id': 'Unique identifier for this cache entry. Alphanumeric, 1-64 chars. Used to retrieve cached KV state.'
• Expand tool descriptions to 100 - 200 chars. Example for insert_prefix: 'Cache a prompt prefix in the radix tree to enable KV-state reuse. Call this once per unique system prompt or template. Idempotent, calling twice with the same kv_cache_id overwrites the previous entry.'
• Document output schemas in tool definitions. Add 'outputSchema' field showing the structure returned by each tool (e.g., insert_prefix returns {status: string}, lookup_prompt returns {total_tokens: int, cached_tokens: int, hit_ratio: float, estimated_ttft_reduction: string, is_full_cache_hit: bool}).
• Add error handling guidance. Return structured errors like {error: 'invalid_prompt', message: 'Prompt must be non-empty and under 10,000 chars', recovery: 'Provide a valid prompt text and retry.'} instead of raw exceptions.
• Add tool annotations (destructiveHint, idempotentHint) to clarify side effects. Mark insert_prefix with idempotentHint=true and destructiveHint=false to signal it is safe to retry.