PromptTuner MCP

// Optimization technique templates for PromptTuner MCP // Based on 2024-2025 prompt engineering best practices from Anthropic, OpenAI, and industry leaders import type { OptimizationTechnique, TargetFormat, TechniqueTemplate, } from '../config/types.js'; const TECHNIQUE_TEMPLATES: Record<OptimizationTechnique, TechniqueTemplate> = { basic: { name: 'basic', description: 'Simple clarity and grammar improvements', systemPrompt: `<task> Improve prompt clarity, fix grammar and spelling, replace vague terms with specifics. </task> <rules> ALWAYS: - Fix spelling and grammar - Replace vague words ("something", "stuff") - Clarify ambiguous references - Add punctuation if missing NEVER: - Add new requirements - Change the core task - Remove important context </rules> Return ONLY the refined prompt.`, }, chainOfThought: { name: 'chainOfThought', description: 'Add step-by-step reasoning', systemPrompt: `<task> Add step-by-step reasoning guidance appropriate for the task. </task> <examples> Math: "Let's calculate step by step." Analysis: "Let's analyze systematically." Debug: "Let's trace through carefully." Planning: "Let's break this into phases." </examples> <rules> ALWAYS: - Add task-appropriate reasoning trigger - Place after the main task NEVER: - Add to simple queries - Use generic triggers </rules> Return ONLY the optimized prompt.`, }, fewShot: { name: 'fewShot', description: 'Add examples (few-shot prompting)', systemPrompt: `<task> Add 2-3 diverse input/output examples to demonstrate the expected behavior. </task> <rules> ALWAYS: - Add 2-3 diverse examples - Show clear input→output pairs - Place examples before the main task - Use <example> tags (Claude) or ### headers (GPT) NEVER: - Add to trivial tasks - Use confusing examples - Use placeholder values like "[example]" </rules> Return ONLY the optimized prompt.`, }, roleBased: { name: 'roleBased', description: 'Add expert persona/role', systemPrompt: `<task> Add an expert role/persona relevant to the task. </task> <examples> Code: "You are a senior software engineer..." Writing: "You are a professional editor..." Analysis: "You are a data analyst..." Teaching: "You are an expert tutor..." </examples> <rules> ALWAYS: - Place role at the start - Use format: "You are [role] with expertise in [domain]" - Keep it concise NEVER: - Use generic roles ("helpful assistant") - Add unnecessary backstory </rules> Return ONLY the optimized prompt.`, }, structured: { name: 'structured', description: 'Add formatting structure (XML/Markdown)', systemPrompt: `<task> Add clear structural formatting appropriate for the target format. </task> <format_guidance> For Claude: - Use XML tags: <context>, <task>, <requirements>, <output> - Keep nesting shallow (max 2 levels) For GPT: - Use ## Headers for sections - Use bullet lists and **bold** - Keep sections scannable </format_guidance> <rules> ALWAYS: - Organize into clear sections - Use lists for multiple items - Match the target format NEVER: - Over-structure simple prompts - Mix XML and Markdown - Nest deeply (max 2 levels) </rules> Return ONLY the structured prompt.`, }, comprehensive: { name: 'comprehensive', description: 'Apply all optimization techniques', systemPrompt: `<task> Apply multiple optimization techniques intelligently to maximize prompt quality. </task> <approach> Consider applying (where beneficial): 1. Expert role (if domain knowledge helps) 2. Clear structure (XML for Claude, Markdown for GPT) 3. Specific, unambiguous language 4. Step-by-step guidance (for complex tasks) 5. Examples (if output format unclear) 6. Explicit constraints (ALWAYS/NEVER) 7. Output format specification </approach> <rules> ALWAYS: - Apply techniques that add value - Maintain original intent - Stay concise NEVER: - Over-engineer simple prompts - Add unnecessary length - Mix formatting styles </rules> Return ONLY the optimized prompt.`, }, }; function getTechniqueTemplate( technique: OptimizationTechnique ): TechniqueTemplate { return TECHNIQUE_TEMPLATES[technique]; } /** Format-specific instruction templates */ const FORMAT_INSTRUCTIONS: Record<Exclude<TargetFormat, 'auto'>, string> = { claude: ` <format_instructions> TARGET FORMAT: Claude (XML structure) Use semantic XML tags to organize content: - <context> for background information - <task> for the main objective - <instructions> for step-by-step guidance - <requirements> for constraints and rules - <examples> for input/output demonstrations - <output_format> for expected response structure </format_instructions>`, gpt: ` <format_instructions> TARGET FORMAT: GPT (Markdown structure) Use Markdown formatting for clarity: - ## Headers for main sections (Context, Task, Requirements) - **Bold** for emphasis on key terms - - Bullet lists for multiple items - 1. Numbered lists for sequential steps - \`code\` for technical terms or values - > Blockquotes for examples or notes </format_instructions>`, json: ` <format_instructions> TARGET FORMAT: JSON output Structure the prompt to request JSON output: - Specify the exact JSON schema expected - Include field descriptions and types - Provide an example of the expected structure - Mention validation requirements if any Example schema specification: { "field_name": "description (type)", "required_field": "must be present", "optional_field?": "may be omitted" } </format_instructions>`, }; function getFormatInstructions(format: TargetFormat): string { if (format === 'auto') return ''; return FORMAT_INSTRUCTIONS[format]; } export function buildRefinementPrompt( originalPrompt: string, technique: OptimizationTechnique, targetFormat: TargetFormat ): string { const template = getTechniqueTemplate(technique); const formatInstructions = getFormatInstructions(targetFormat); return `${template.systemPrompt} ${formatInstructions} ORIGINAL PROMPT: ${originalPrompt}`; }