Server has 8 tools with complete JSON schemas and descriptions. Naming follows verb_noun convention (get_*, update_*, save_*). However, descriptions are verbose and written in Chinese with mixed English, making them harder for LLMs to parse reliably. Parameter descriptions exist but lack clarity on constraints, ranges, and error conditions. Output schemas are not documented. Error handling guidance is minimal. Tool composition is reasonable but some tools conflate multiple concerns (e.g., get_cards returns 7 card types in one call rather than allowing filtering).

• Rewrite all descriptions in English, 50-200 characters, following the pattern: WHAT (verb + object) + WHEN (use case) + RETURN (key fields). Example: 'Fetch the 7 identity cards (preferences, goals, etc.). Call this at session start or after manual dashboard updates. Returns card_type, version, content, locked status.'
• Document output schemas for all tools. For get_cards, specify: returns array of {card_type: string, version: int, content: string, locked: bool, last_updated: ISO8601, char_count: int}. For get_conversations, specify: returns array of {conversation_id, summary, created_at, turn_count}.
• Add parameter descriptions for all optional fields. E.g., get_archive.include_raw: 'If true, return full raw_text (counts toward storage tier). Default false saves tokens.' get_conversations.limit: 'Number of recent conversations to return (1-100, default 30).'
• Add error handling guidance to update_card and save_conversation. E.g., 'If content exceeds 2000 chars, truncate and retry. If locked=true, inform user that only dashboard can unlock.' For save_conversation: 'If raw_text exceeds tier limit, omit it and save summary only.'
• Refactor get_cards to accept optional card_type parameter (enum of 7 types) to filter results. Reduces token waste when only one card is needed. Alternatively, create get_card(card_type) for single-card access.