PromptTuner MCP

// Detect Format Tool - LLM-powered format detection import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'; import type { FormatDetectionResponse } from '../config/types.js'; import { createErrorResponse, createSuccessResponse, ErrorCode, } from '../lib/errors.js'; import { executeLLMWithJsonResponse } from '../lib/tool-helpers.js'; import { validatePrompt } from '../lib/validation.js'; import { DetectFormatInputSchema, DetectFormatOutputSchema, } from '../schemas/index.js'; import { FormatDetectionResponseSchema } from '../schemas/llm-responses.js'; const FORMAT_DETECTION_PROMPT = `You are an expert at detecting AI prompt formats. Analyze the prompt and determine its target format. <formats> 1. **Claude XML**: Uses <context>, <task>, <requirements>, <output_format> tags 2. **GPT Markdown**: Uses ## headers, **bold**, bullet lists, numbered steps 3. **JSON**: Structured data format with schemas and key-value patterns 4. **Auto/Generic**: No specific format detected </formats> <detection_criteria> Claude XML indicators: - Presence of XML tags like <context>, <task>, <instructions> - Structured semantic sections in angle brackets - Anthropic-style formatting GPT Markdown indicators: - ## Headers for sections - **Bold** emphasis - Bullet lists with - or * - Markdown syntax JSON indicators: - Schema definitions - Key-value pair structures - Curly braces and quotes - Data extraction focus </detection_criteria> **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 Example valid response: { "detectedFormat": "gpt", "confidence": 85, "recommendation": "This prompt uses Markdown headers (##), bold emphasis (**), and bullet lists, indicating GPT format. Consider adding more specific examples." } Required JSON schema: { "detectedFormat": "claude" | "gpt" | "json" | "auto", "confidence": number (0-100), "recommendation": string }`; /** * Registers the detect_format tool with the MCP server. * Uses external LLM to detect format. */ export function registerDetectFormatTool(server: McpServer): void { server.registerTool( 'detect_format', { title: 'Detect Format', description: 'Identify if prompt targets Claude XML, GPT Markdown, or JSON schema using AI analysis. Returns confidence score and recommendations.', inputSchema: DetectFormatInputSchema, outputSchema: DetectFormatOutputSchema, annotations: { readOnlyHint: true, idempotentHint: false, openWorldHint: false, }, }, async ({ prompt, }): Promise< | ReturnType<typeof createSuccessResponse> | ReturnType<typeof createErrorResponse> > => { try { const validatedPrompt = validatePrompt(prompt); const detectionPrompt = `${FORMAT_DETECTION_PROMPT}\n\nPROMPT TO ANALYZE:\n${validatedPrompt}`; const parsed = await executeLLMWithJsonResponse<FormatDetectionResponse>( detectionPrompt, (value) => FormatDetectionResponseSchema.parse(value), ErrorCode.E_LLM_FAILED, 'detect_format', { maxTokens: 500 } ); const output = `# Format Detection\n\n**Detected Format**: ${parsed.detectedFormat}\n**Confidence**: ${parsed.confidence}%\n\n**Recommendation**:\n${parsed.recommendation}`; return createSuccessResponse(output, { ok: true, detectedFormat: parsed.detectedFormat, confidence: parsed.confidence, recommendation: parsed.recommendation, }); } catch (error) { return createErrorResponse(error, ErrorCode.E_LLM_FAILED, prompt); } } ); }