Archivist has 4 tools with clear names and generally good descriptions. All tools have input schemas with types and descriptions. However, output schemas are not documented, parameter descriptions lack depth in some cases, and error handling is minimal. Tool naming follows verb_noun convention (search, save_memory, recall, recent_memories). Descriptions are concise (10-60 chars) but lack detail on what happens internally, when to use each tool vs others, and what the LLM should expect from responses. Schemas are present with types and defaults, but lack constraints (enums, ranges, patterns). No tool annotations (readOnlyHint/destructiveHint), no documented output structure, and no error recovery guidance.

• Add explicit output schema documentation for each tool. Example: 'search returns {results: [{chunk: string, similarity: number, doc_id: string}], total: integer}'. This tells the LLM what to expect and enables chaining.
• Expand tool descriptions to 80-150 characters. Include WHEN to use (e.g., 'Call search when the user asks a factual question about ingested documents. Use recall for context about the user themselves.'). Distinguish search (documents) from recall (memories) in descriptions.
• Add tool annotations in MCP schema: search with readOnlyHint=true, save_memory with destructiveHint=false (persist operation, not destructive) and idempotentHint=true, recall with readOnlyHint=true, recent_memories with readOnlyHint=true.
• Document parameter constraints: 'limit: integer 1-50 (default 5 for search, default 10 for recent_memories, default 5 for recall)'. Add regex pattern or enum for tags if there are restricted tag categories (e.g., ['preference', 'project', 'tool', 'user']).
• Add error handling guidance to search and recall: 'If no documents are indexed, search returns an empty results array. Call GET /collections to verify documents have been ingested.' For recall: 'If no memories match the query, consider calling recent_memories to browse available context.'
• Document pagination for search and recall: specify whether responses include total_count, next_cursor, or offset. If result sets can be large, clarify truncation behavior.
• Include sample outputs in resource guides (guide://search, guide://memory). Example: 'A search result is: {chunk: "...", similarity: 0.92, doc_id: "..."}'. This helps the LLM understand structure without needing to infer from the first call.