Single tool 'vroid_studio' has a descriptive name but suffers from critical design flaws. The tool conflates 15+ distinct operations (status, launch, focus, screenshot, list_shortcuts, list_archetypes, run_archetype, run_template, open_project, save_project, save_project_as, open_and_export, list_templates, quick_gal_export, list_outputs, session_status) into one parameter-driven dispatcher. This violates the single-responsibility principle and forces LLMs to reason about which operation to invoke rather than selecting a tool by intent. The description is adequate (97 chars) but generic. Parameters lack enums for the 'operation' field despite 16 known values, inviting hallucinated operations. Most parameters have descriptions but lack type constraints (e.g., 'archetype_id' defaults to 'quick_gal' with no enum of valid archetypes). Output schema is not documented, callers cannot predict response structure. Error handling is absent from the visible code.

• Split vroid_studio into 16 separate tools: launch_vroid_studio(), focus_vroid_studio(), screenshot_vroid_studio(), list_archetypes(), run_archetype(archetype_id), run_template(template_name), open_project(project_path), save_project(project_path), save_project_as(project_path, save_path), open_and_export(project_path, output_name), list_templates(), quick_gal_export(output_name, pick_sample), list_outputs(), get_vroid_status(), get_session_status(session_id). Each tool has one clear responsibility and LLMs can select by intent.
• Add enum constraints to all parameters with known finite sets. For 'archetype_id', enumerate all 55 valid values. For 'template_name', enumerate all templates. For 'operation' (if kept), declare the 16 valid operations as an enum.
• Document output schemas for each operation. Use a discriminated union pattern: {operation: 'status', success: bool, running: bool, handle: string, outputs: [string], archetype_count: int} vs {operation: 'run_archetype', success: bool, session_id: string, steps_completed: int, verification_screenshots: [string]}. Or split into separate tools with explicit return types.
• Add error handling with recovery guidance. Wrap operations in try-catch and return structured errors: {success: false, error: 'VRoid Studio not running', recovery: 'Call launch_vroid_studio() first'} or {success: false, error: 'Invalid archetype_id: quick_gal_invalid', recovery: 'Call list_archetypes() to see valid options: [quick_gal, full_body_export, ...]'}.
• Provide a discovery tool list_archetypes() that returns {archetypes: [{id: string, name: string, description: string}]} so agents can explore available options before calling run_archetype(). Similarly, list_templates() should return available templates with descriptions.