Graft provides 5 well-named, read-only tools with clear descriptions (avg 85 chars) and structured input schemas. All tools follow verb_noun naming (graft_map, graft_context, graft_search, graft_impact, graft_summary). Schemas include type definitions and descriptions for all parameters. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal (path validation only, no recovery guidance). No tool annotations (readOnlyHint, idempotentHint). Parameter descriptions lack format/constraint details (e.g., 'query' maxLength=500 not mentioned in description text). Missing pagination guidance for potentially large result sets.

• Document output schemas for all tools. Example: graft_map returns {text: string, tokenCount: number}; graft_search returns {matches: [{kind, name, file, line}], total: number}. This enables LLMs to chain tools and extract fields correctly.
• Add tool annotations to server registration: readOnlyHint=true for all tools (they never modify state). This signals to LLMs that retries are safe and no confirmation is needed.
• Enhance parameter descriptions with constraint details. Change 'File or symbol for personalization' to 'File or symbol for personalization (max 500 chars, relative to project root)'. Repeat maxLength/minimum/maximum constraints in description text.
• Improve error messages with recovery guidance. Change 'path is outside project root' to 'Path is outside project root. Use relative paths from the project root, e.g., src/index.ts'.
• Document graft_summary output. Clarify what 'project statistics' means (file count, line count, language breakdown?) and what 'key files' are (entry points, most-imported files?). Add example output structure.