Two tools with basic schemas and descriptions, but significant gaps in parameter documentation, output schema clarity, and error handling. optimize_prompt has a reasonable description (127 chars), but score_prompt is minimal (95 chars). Neither tool documents return structure or error cases. Parameters lack depth: 'mode' enum is present but 'projectRoot' has no guidance on format or when to use it. No error recovery guidance. Tool composition is reasonable (single responsibility), but output handling is opaque, both return text blobs rather than structured objects. Missing: per-parameter descriptions for all inputs, documented output schemas, error classification, and recovery hints.

• Add explicit output schema documentation to both tools. E.g., optimize_prompt returns {optimized: string, score_breakdown: {clarity, specificity, completeness, efficiency}, tokens_saved: number, cost_savings: string}.
• Expand score_prompt description to 150+ chars: 'Analyzes a prompt across 4 dimensions (clarity, specificity, completeness, efficiency). Returns numeric scores (0-100 per dimension) and a list of rule violations with severity levels. Use this to diagnose prompt weaknesses before optimize_prompt.'
• Add description to projectRoot parameter: 'Optional. Absolute or relative path to project root. Used for contextual rewriting, if provided, optimization rules consider project structure and naming conventions. Omit for generic optimization.'
• Implement structured error responses. E.g., if raw_prompt is empty, return {error: 'empty_input', message: 'Prompt cannot be empty', suggestion: 'Provide a non-empty prompt text'}. Guide LLM to retry with valid input.
• Return score_prompt result as structured JSON object (not stringified), with fields: {total_score: number, breakdown: {clarity, specificity, completeness, efficiency}, violations: [{id, dimension, severity, message}]}. Enables reliable downstream parsing.