PromptTuner MCP

import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'; import type { OptimizeResponse } from '../config/types.js'; import { createErrorResponse, createSuccessResponse, ErrorCode, } from '../lib/errors.js'; import { resolveFormat } from '../lib/prompt-analysis.js'; import { executeLLMWithJsonResponse } from '../lib/tool-helpers.js'; import { validateFormat, validatePrompt, validateTechniques, } from '../lib/validation.js'; import { OptimizePromptInputSchema, OptimizePromptOutputSchema, } from '../schemas/index.js'; import { OptimizeResponseSchema } from '../schemas/llm-responses.js'; const OPTIMIZE_SYSTEM_PROMPT = `You are an expert prompt engineer. <task> Apply the requested optimization techniques to improve the prompt. </task> <techniques> Available techniques: 1. basic - Fix grammar, spelling, clarity 2. chainOfThought - Add step-by-step reasoning 3. fewShot - Add 2-3 diverse examples 4. roleBased - Add expert persona 5. structured - Add XML (Claude) or Markdown (GPT) 6. comprehensive - Apply all intelligently </techniques> <rules> ALWAYS: - Apply techniques in requested order - Build on previous refinements - Maintain original intent - Provide before/after scores - Return plain text (no markdown code blocks) NEVER: - Over-engineer simple prompts - Wrap output in triple backticks - Change the core task </rules> <scoring> Score before and after (0-100): - Clarity - Specificity - Completeness - Structure - Effectiveness - Overall (weighted average) </scoring> <output> **CRITICAL: Your response MUST be valid, parseable JSON only. No markdown, no code blocks, no explanatory text.** 1. Start your response with { (opening brace) 2. End your response with } (closing brace) 3. Use proper JSON syntax: double quotes for strings, no trailing commas 4. All required fields MUST be present 5. Do NOT wrap in \`\`\`json code blocks 6. Escape special characters in strings: \\n for newlines, \\" for quotes, \\\\ for backslashes Example valid response: { "optimized": "You are a senior software engineer. Analyze the code step-by-step:\n\n1. Identify bugs\n2. Suggest improvements\n3. Rate code quality (1-10)", "techniquesApplied": ["basic", "roleBased", "structured"], "improvements": [ "Fixed grammar and spelling errors", "Added expert role context (senior engineer)", "Structured output with numbered steps" ], "beforeScore": { "clarity": 60, "specificity": 55, "completeness": 50, "structure": 40, "effectiveness": 52, "overall": 51 }, "afterScore": { "clarity": 90, "specificity": 85, "completeness": 88, "structure": 95, "effectiveness": 89, "overall": 89 } } </output> <schema> { "optimized": string (the fully optimized prompt with proper escaping), "techniquesApplied": string[] (array of technique names), "improvements": string[] (array of changes made), "beforeScore": { "clarity": number (0-100), "specificity": number (0-100), "completeness": number (0-100), "structure": number (0-100), "effectiveness": number (0-100), "overall": number (0-100) }, "afterScore": { "clarity": number (0-100), "specificity": number (0-100), "completeness": number (0-100), "structure": number (0-100), "effectiveness": number (0-100), "overall": number (0-100) } } </schema>`; function formatScoreSection( label: string, score: OptimizeResponse['beforeScore'] ): string { return [ `## ${label}: ${score.overall}/100`, `- Clarity: ${score.clarity}`, `- Specificity: ${score.specificity}`, `- Completeness: ${score.completeness}`, `- Structure: ${score.structure}`, `- Effectiveness: ${score.effectiveness}`, ].join('\n'); } function formatOptimizeOutput(optimizationResult: OptimizeResponse): string { return [ `# Optimization Results`, ``, formatScoreSection('Before Score', optimizationResult.beforeScore), ``, formatScoreSection('After Score', optimizationResult.afterScore), ``, `## Techniques Applied`, optimizationResult.techniquesApplied .map((technique) => `- ${technique}`) .join('\n'), ``, `## Improvements`, optimizationResult.improvements .map((improvement) => `- ${improvement}`) .join('\n'), ``, `## Optimized Prompt`, '```', optimizationResult.optimized, '```', ].join('\n'); } export function registerOptimizePromptTool(server: McpServer): void { server.registerTool( 'optimize_prompt', { title: 'Optimize Prompt', description: 'Apply multiple optimization techniques using AI (e.g., ["basic", "roleBased", "structured"]). Returns before/after scores and improvements.', inputSchema: OptimizePromptInputSchema, outputSchema: OptimizePromptOutputSchema, annotations: { readOnlyHint: true, idempotentHint: false, openWorldHint: false, }, }, async ({ prompt, techniques, targetFormat, }): Promise< | ReturnType<typeof createSuccessResponse> | ReturnType<typeof createErrorResponse> > => { try { const validatedPrompt = validatePrompt(prompt); const validatedTechniques = validateTechniques(techniques); const validatedFormat = validateFormat(targetFormat); const resolvedFormat = resolveFormat(validatedFormat, validatedPrompt); const optimizePrompt = `${OPTIMIZE_SYSTEM_PROMPT}\n\nTarget Format: ${resolvedFormat}\nTechniques to apply: ${validatedTechniques.join(', ')}\n\nORIGINAL PROMPT:\n${validatedPrompt}`; const optimizationResult = await executeLLMWithJsonResponse<OptimizeResponse>( optimizePrompt, (value) => OptimizeResponseSchema.parse(value), ErrorCode.E_LLM_FAILED, 'optimize_prompt', { maxTokens: 3000, timeoutMs: 60000 } ); const scoreDiff = optimizationResult.afterScore.overall - optimizationResult.beforeScore.overall; if (scoreDiff > 0) { optimizationResult.improvements.push( `Overall score improved by ${scoreDiff} points` ); } const output = formatOptimizeOutput(optimizationResult); return createSuccessResponse(output, { ok: true, original: validatedPrompt, optimized: optimizationResult.optimized, techniquesApplied: optimizationResult.techniquesApplied, targetFormat: resolvedFormat, beforeScore: optimizationResult.beforeScore, afterScore: optimizationResult.afterScore, improvements: optimizationResult.improvements, usedFallback: false, }); } catch (error) { return createErrorResponse(error, ErrorCode.E_LLM_FAILED, prompt); } } ); }