arxiv-mcp demonstrates moderate definition quality with significant strengths in naming and structure but notable gaps in parameter descriptions and output schema documentation. 33 tools show consistent verb-noun naming (search_papers, fetch_full_text, list_categories). Tool descriptions are present but vary in completeness and LLM-optimization, many are overly technical or include implementation details (e.g., 'HTML→Markdown with PDF fallback', 'Jina Reader full text') rather than user-intent framing. Parameter schemas are visible but lack comprehensive descriptions for many parameters (e.g., 'preferred_html' is described minimally, 'format' enums lack rationale). Output schemas are not documented, callers must infer return types from tool names alone. Error handling is implicit; no recovery guidance visible in descriptions. Security consideration: no apparent secret injection for API keys (Semantic Scholar, Jina Reader credentials likely needed). Composition is strong, tools chain logically (search→get_details→fetch_full_text→ingest). The epistemic analysis and depot RAG features are sophisticated but lack clear invocation guidance for users unfamiliar with the domain.

• Document output schema for every tool. Use JSON Schema or structured text in the tool description. E.g., 'Returns: {papers: [{id, title, authors, abstract, url}], total_count, next_cursor}'. This enables agents to plan follow-up calls and extract needed fields.
• Standardize naming conventions. Use snake_case for all tools (rename getPaper → get_paper, getContent → get_content, etc.). This reduces LLM confusion when multiple naming styles are present.
• Expand parameter descriptions to include constraints and examples. E.g., 'sort_by: relevance (default, matches query keywords) | submitted (newest first) | updated (recently revised). Default: relevance.' Enums must explain when to use each option.
• Add recovery guidance to descriptions of tools that may fail. E.g., fetch_full_text: 'If HTML extraction times out, returns a hint to retry with prefer_html=false. If PDF is unavailable, returns error with reference IDs for manual lookup.'
• Refactor specialized descriptions for clarity. Define domain terms inline: 'analyze_paper_epistemics: Classify evidence type (simulation | theoretical | empirical | meta-analysis) and flag claims needing bench validation, telescope observation, or formal proof-checking.' This grounds domain jargon.
• Add 'When to call' hints in descriptions. E.g., 'list_category_latest: Call this first to discover recent papers in a topic before deep analysis.' This accelerates agent planning.
• Consolidate backward-compatibility aliases. Remove fetch_anthropic_post and list_anthropic_posts; use fetch_lab_post and list_lab_posts with source="anthropic". This eliminates duplicate tool confusion.