ao3-mcp is a specialized single-purpose server with 5 well-motivated tools focused on AO3 searching and reading delegation. Tool names follow verb_noun conventions (search_works, find_tags, get_work, read_works, get_work_text). Descriptions are detailed and rich, contextualizing the READING DELEGATION pattern, the core innovation, within each tool's docstring. Parameters are largely well-typed and described. However, there are measurable gaps: (1) some parameters lack explicit enums where they should have them (e.g., 'rating', 'categories', 'sort_by'), (2) no explicit output schemas are documented, (3) error handling guidance is absent from tool docstrings, (4) the 'query' parameter in search_works is dense and relies heavily on inline documentation rather than structured constraints. The server compensates with domain-specific context (AO3-specific operators, wildcard advice) embedded in descriptions, which aids LLM decision-making. Per-tool average: 62.

• Add formal JSON Schema output definitions for all 5 tools. For search_works, document the structure of each work object (work_id, title, authors, fandoms, rating, category, words, chapters, complete, kudos, hits, comments, bookmarks, summary, tags, relationships, updated). For read_works, specify the structure of the 'structured digest' (e.g., {plot: string, characters: [string], prose_quality: string, content_notes: [string]}).
• Convert constrained string parameters to enums: search_works.rating → enum {not_rated, general, teen, mature, explicit}; search_works.categories → enum {gen, f/m, m/m, other, f/f, multi}; search_works.sort_by → enum {relevance, kudos, hits, comments, bookmarks, words, date_updated, date_posted}; find_tags.kind → enum {tag, fandom, relationship, character}.
• Add error handling guidance to each tool. Example for read_works: 'May return empty text if the Gemini reader refuses the content (usually explicit/extreme). This is intermittent; retry or use get_work_text instead.' Example for search_works: 'Rate-limited by AO3 (throttle enforced server-side). If you hit the limit, wait 1+ minute before retrying.'
• Simplify pagination in search_works by adding a 'limit' parameter (default 20, max 100) in addition to or instead of 'pages'. Document the total result count in the response so the LLM knows if pagination is needed.
• Consider extracting common AO3 query operators into separate optional parameters: min_kudos, max_kudos, min_hits, max_hits, exclude_term, one_shot_only. This reduces the cognitive load on the LLM to construct complex query strings and makes the tool self-documenting.