Claude Atlas has 12 tools with complete input schemas and descriptions. Naming follows verb_noun convention (create_project, get_project, archive_project). Descriptions are present and substantive (100-250 chars typical), explaining what each tool does. However, parameter descriptions lack detail on constraints, formats, and valid ranges. Output schemas are not documented, responses are formatted as JSON strings rather than structured objects. Error handling returns raw exception messages instead of actionable recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Missing pagination guidance for list_projects despite accepting a limit parameter.

• Document output schemas for all tools. Define the structure of returned objects (e.g., project object with fields: key, name, description, state, dueDate, owner, members, tags). Return structured JSON objects, not stringified JSON.
• Enhance error handling with categorization and recovery guidance. Instead of 'Error: RuntimeError: Goal not found', return '{"error": "goal_not_found", "message": "Goal GOAL-42 not found in this site. Try list_goals() to see available goals.", "retryable": false}'.
• Add constraint details to parameter descriptions. For 'goal' parameter: 'Goal key (format: GOAL-\d+, e.g. GOAL-42) or goal ARI (ari:cloud:atlas::goal/...). Resolved by key within the current site.' For 'tag': 'Tag name (1-100 characters, alphanumeric + spaces/hyphens).'
• Declare tool annotations in FastMCP registration. Mark archive_project, unarchive_project, create_project, create_project_update with destructiveHint=true. Mark get_project, get_projects, list_projects, get_project_updates with readOnlyHint=true.
• Add pagination guidance to list_projects. Document that limit defaults to 50, explain how to fetch additional results (offset-based or cursor-based), and state the maximum result set size to prevent context window exhaustion.