Latitude MCP Server

/** * Prompting Techniques Documentation * Comprehensive guides for each technique pattern */ // ============================================================================ // TECHNIQUE-REACT // ============================================================================ export const DOCS_TECHNIQUE_REACT = `# ReAct Pattern > Reasoning + Acting in a loop ## Overview ReAct (Reasoning and Acting) combines chain-of-thought reasoning with action-taking. The AI thinks through problems step by step while using tools to gather information. --- ## Quick Reference \`\`\` Think → Act → Observe → Repeat \`\`\` --- ## The ReAct Loop 1. **Thought**: Reason about what to do 2. **Action**: Call a tool or take an action 3. **Observation**: Process the result 4. **Repeat**: Until goal is achieved --- ## Basic ReAct Agent \`\`\`promptl --- provider: OpenAI model: gpt-4o type: agent maxSteps: 10 tools: - latitude/search - latitude/extract --- <system> You are a research assistant that follows the ReAct pattern. For each step: 1. THOUGHT: Explain your reasoning 2. ACTION: Use a tool if needed 3. OBSERVATION: Analyze the result 4. Repeat until you have the answer Always show your thinking process. </system> <user> {{ question }} </user> \`\`\` --- ## Structured ReAct \`\`\`promptl --- provider: OpenAI model: gpt-4o type: agent tools: - search_database: description: Search internal database parameters: type: object properties: query: type: string required: [query] - get_details: description: Get details for an item parameters: type: object properties: id: type: string required: [id] --- <system> Follow the ReAct pattern strictly: **THOUGHT**: [What do I need to do? What information do I need?] **ACTION**: [Which tool to use and why] **OBSERVATION**: [What did I learn from the result?] Continue until you can provide a complete answer. </system> <user> {{ task }} </user> \`\`\` --- ## ReAct vs Standard Agent | Aspect | Standard Agent | ReAct Agent | |--------|---------------|-------------| | Transparency | Low | High (shows reasoning) | | Debugging | Harder | Easier (visible thoughts) | | Token usage | Lower | Higher | | Accuracy | Good | Better (explicit reasoning) | --- ## When to Use ReAct | Scenario | Use ReAct? | |----------|------------| | Research tasks | Yes ✅ | | Multi-step problems | Yes ✅ | | Debugging needed | Yes ✅ | | Simple lookups | No (overkill) | | High volume | No (too many tokens) | --- ## Best Practices - ✅ **DO**: Require explicit THOUGHT/ACTION/OBSERVATION format - ✅ **DO**: Set appropriate maxSteps limit - ✅ **DO**: Use for complex, multi-step tasks - ❌ **DON'T**: Use for simple tasks - ❌ **DON'T**: Skip the observation step --- ## Next Steps - **Agents** → \`docs({ action: "get", topic: "agents" })\` - **Chain of Thought** → \`docs({ action: "get", topic: "technique-cot" })\` `; // ============================================================================ // TECHNIQUE-SELF-CONSISTENCY // ============================================================================ export const DOCS_TECHNIQUE_SELF_CONSISTENCY = `# Self-Consistency > Multiple answers, majority vote ## Overview Self-consistency generates multiple reasoning paths and selects the most common answer. Improves accuracy on reasoning tasks by 5-10%. --- ## Quick Reference \`\`\` Generate N solutions → Vote → Select majority \`\`\` --- ## Basic Self-Consistency \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="answer1" temperature={{ 0.7 }}> {{ problem }} Think step by step and give your final answer. </step> <step as="answer2" temperature={{ 0.7 }}> {{ problem }} Think step by step and give your final answer. </step> <step as="answer3" temperature={{ 0.7 }}> {{ problem }} Think step by step and give your final answer. </step> <step> I generated three solutions: 1. {{ answer1 }} 2. {{ answer2 }} 3. {{ answer3 }} What is the most common answer? Provide only that answer. </step> \`\`\` --- ## Structured Self-Consistency \`\`\`promptl --- provider: OpenAI model: gpt-4o schema: type: object properties: final_answer: type: string confidence: type: number reasoning: type: string required: [final_answer, confidence] --- <step isolated as="sol1" temperature={{ 0.8 }}> {{ problem }} Solve step by step. </step> <step isolated as="sol2" temperature={{ 0.8 }}> {{ problem }} Solve step by step. </step> <step isolated as="sol3" temperature={{ 0.8 }}> {{ problem }} Solve step by step. </step> <step> Compare these solutions: 1. {{ sol1 }} 2. {{ sol2 }} 3. {{ sol3 }} Determine the consensus answer. - If all agree: confidence = 1.0 - If 2/3 agree: confidence = 0.7 - If all different: pick best, confidence = 0.3 </step> \`\`\` --- ## When to Use | Task | Self-Consistency? | |------|-------------------| | Math problems | Yes ✅ | | Logic puzzles | Yes ✅ | | Code generation | Yes ✅ | | Creative writing | No | | Simple questions | No (overkill) | --- ## Number of Samples | Samples | Accuracy Boost | Cost | |---------|----------------|------| | 3 | Moderate | 3x | | 5 | Good | 5x | | 10+ | Diminishing returns | 10x+ | **Recommendation**: Start with 3, increase if needed. --- ## Best Practices - ✅ **DO**: Use temperature > 0.5 for diversity - ✅ **DO**: Use isolated steps to prevent contamination - ✅ **DO**: Use for high-stakes reasoning - ❌ **DON'T**: Use for subjective questions - ❌ **DON'T**: Use when cost is a concern --- ## Next Steps - **Chain of Thought** → \`docs({ action: "get", topic: "technique-cot" })\` - **Tree of Thoughts** → \`docs({ action: "get", topic: "technique-tot" })\` `; // ============================================================================ // TECHNIQUE-CONSTITUTIONAL // ============================================================================ export const DOCS_TECHNIQUE_CONSTITUTIONAL = `# Constitutional AI > Principle-based self-critique ## Overview Constitutional AI defines principles the AI must follow, then critiques its own outputs against those principles. Creates safer, more aligned responses. --- ## Quick Reference \`\`\` Generate → Critique against principles → Revise \`\`\` --- ## Basic Constitutional Pattern \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> CONSTITUTIONAL PRINCIPLES: 1. Be helpful, harmless, and honest 2. Never provide harmful information 3. Respect user privacy 4. Acknowledge limitations 5. Recommend professionals when appropriate Always check your response against these principles before answering. </system> <user> {{ question }} </user> \`\`\` --- ## Self-Critique Pattern \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="draft"> {{ request }} </step> <step as="critique"> Review this response against our principles: PRINCIPLES: - Be helpful and accurate - Never provide harmful information - Be respectful and inclusive - Acknowledge uncertainty Response to review: {{ draft }} Issues found (list each): </step> <step> {{ if critique contains issues }} Revise the response to address: {{ critique }} Original: {{ draft }} {{ else }} {{ draft }} {{ endif }} </step> \`\`\` --- ## Domain-Specific Principles ### Medical Assistant \`\`\`promptl <system> MEDICAL PRINCIPLES: 1. Never diagnose - suggest consulting a doctor 2. Provide general health information only 3. Recommend emergency services for urgent symptoms 4. Cite reputable sources (CDC, WHO, NIH) 5. Acknowledge not being a medical professional </system> \`\`\` ### Financial Assistant \`\`\`promptl <system> FINANCIAL PRINCIPLES: 1. Never provide specific investment advice 2. Always recommend consulting a financial advisor 3. Explain risks alongside opportunities 4. Use disclaimers for financial information 5. Don't guarantee returns </system> \`\`\` --- ## When to Use | Scenario | Constitutional AI? | |----------|-------------------| | Sensitive topics | Yes ✅ | | User safety critical | Yes ✅ | | Regulated domains | Yes ✅ | | General chat | Optional | | Internal tools | Optional | --- ## Best Practices - ✅ **DO**: Define clear, specific principles - ✅ **DO**: Include self-critique step for sensitive tasks - ✅ **DO**: Update principles based on failures - ❌ **DON'T**: Use too many principles (5-10 max) - ❌ **DON'T**: Make principles too vague --- ## Next Steps - **Safety guide** → \`docs({ action: "get", topic: "guide-safety" })\` - **Iterative refinement** → \`docs({ action: "get", topic: "technique-iterative" })\` `; // ============================================================================ // TECHNIQUE-SOCRATIC // ============================================================================ export const DOCS_TECHNIQUE_SOCRATIC = `# Socratic Questioning > Guide through questions, not answers ## Overview Instead of giving direct answers, guide users to discover answers themselves through thoughtful questions. Excellent for education and coaching. --- ## Quick Reference \`\`\` Ask clarifying questions → Guide thinking → Let user discover \`\`\` --- ## Basic Socratic Tutor \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> You are a Socratic tutor. Never give direct answers. Instead: 1. Ask clarifying questions 2. Break complex problems into parts 3. Guide toward discovery 4. Celebrate insights 5. Offer hints only when stuck Example: User: "What is 15% of 80?" You: "Let's think about this. What does 'percent' mean? And how might you find a part of a whole number?" </system> <user> {{ question }} </user> \`\`\` --- ## Structured Socratic Flow \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> Follow the Socratic method: 1. UNDERSTAND: "What do you already know about this?" 2. PROBE: "What assumptions are you making?" 3. CHALLENGE: "Is there another way to look at this?" 4. IMPLICATIONS: "What follows from that reasoning?" 5. REFLECT: "What have you learned?" Never reveal the answer. Guide the student to find it. </system> {{ for msg in history }} <{{ msg.role }}>{{ msg.content }}</{{ msg.role }}> {{ endfor }} <user>{{ message }}</user> \`\`\` --- ## With Difficulty Levels \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> You are a {{ subject }} tutor using Socratic method. Difficulty: {{ difficulty }} {{ if difficulty == "beginner" }} - Ask simple yes/no questions - Provide more scaffolding - Celebrate small wins {{ elseif difficulty == "intermediate" }} - Ask open-ended questions - Provide moderate hints {{ else }} - Ask challenging questions - Minimal hints - Encourage deeper exploration {{ endif }} </system> <user>{{ question }}</user> \`\`\` --- ## When to Use | Scenario | Socratic? | |----------|-----------| | Teaching/tutoring | Yes ✅ | | Coaching | Yes ✅ | | Problem-solving training | Yes ✅ | | Quick answers needed | No | | Technical support | No | --- ## Best Practices - ✅ **DO**: Be patient and encouraging - ✅ **DO**: Break problems into smaller parts - ✅ **DO**: Adapt to user's level - ❌ **DON'T**: Give direct answers when asked - ❌ **DON'T**: Ask leading questions with obvious answers --- ## Next Steps - **Role prompting** → \`docs({ action: "get", topic: "technique-role" })\` - **Chain of Thought** → \`docs({ action: "get", topic: "technique-cot" })\` `; // ============================================================================ // TECHNIQUE-META // ============================================================================ export const DOCS_TECHNIQUE_META = `# Meta-Prompting > AI generates or improves prompts ## Overview Use AI to generate, analyze, or improve prompts. The AI becomes a prompt engineer, creating better prompts than humans might write. --- ## Quick Reference \`\`\` Describe goal → AI generates prompt → Refine → Use \`\`\` --- ## Prompt Generator \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> You are an expert prompt engineer. Generate PromptL prompts based on user requirements. Include: 1. Appropriate provider/model 2. Temperature settings 3. System message with role 4. User message with variables 5. Output schema if structured output needed </system> <user> Create a prompt for: {{ use_case }} Requirements: {{ requirements }} </user> \`\`\` --- ## Prompt Improver \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> You are a prompt optimization expert. Analyze and improve prompts. Evaluate: 1. Clarity of instructions 2. Appropriate constraints 3. Good examples (if few-shot) 4. Output format specification 5. Edge case handling Provide the improved version with explanations. </system> <user> Improve this prompt: \`\`\` {{ original_prompt }} \`\`\` Goal: {{ goal }} Issues: {{ issues }} </user> \`\`\` --- ## Prompt Analyzer \`\`\`promptl --- provider: OpenAI model: gpt-4o schema: type: object properties: score: type: integer description: 1-10 quality score strengths: type: array items: type: string weaknesses: type: array items: type: string suggestions: type: array items: type: string required: [score, strengths, weaknesses, suggestions] --- Analyze this prompt's quality: \`\`\` {{ prompt_to_analyze }} \`\`\` Intended use: {{ intended_use }} \`\`\` --- ## Auto-Improve Loop \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="analysis"> Analyze this prompt for weaknesses: {{ prompt }} </step> <step as="improved"> Based on this analysis: {{ analysis }} Generate an improved version. </step> <step> IMPROVED PROMPT: {{ improved }} CHANGES MADE: [explain what changed and why] </step> \`\`\` --- ## When to Use | Scenario | Meta-Prompting? | |----------|-----------------| | Creating new prompts | Yes ✅ | | Debugging poor results | Yes ✅ | | Prompt library creation | Yes ✅ | | Quick one-off task | No (overkill) | --- ## Best Practices - ✅ **DO**: Be specific about requirements - ✅ **DO**: Test generated prompts - ✅ **DO**: Iterate with feedback - ❌ **DON'T**: Blindly trust generated prompts - ❌ **DON'T**: Over-complicate simple needs --- ## Next Steps - **Iterative refinement** → \`docs({ action: "get", topic: "technique-iterative" })\` - **Testing** → \`docs({ action: "get", topic: "guide-testing" })\` `; // ============================================================================ // TECHNIQUE-ITERATIVE // ============================================================================ export const DOCS_TECHNIQUE_ITERATIVE = `# Iterative Refinement > Generate → Critique → Improve loop ## Overview Iterative refinement improves output through multiple passes. Generate a draft, critique it, then refine based on the critique. --- ## Quick Reference \`\`\` Draft → Critique → Revise → Repeat \`\`\` --- ## Basic Iteration \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="draft"> Write a {{ content_type }} about {{ topic }}. </step> <step as="critique"> Review this draft and list specific improvements: {{ draft }} Focus on: - Clarity - Accuracy - Engagement - Missing information </step> <step> Revise the draft based on this feedback: {{ critique }} Original: {{ draft }} </step> \`\`\` --- ## Multi-Round Iteration \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="v1"> {{ task }} </step> <step as="critique1"> Critique v1: {{ v1 }} List 3 improvements. </step> <step as="v2"> Improve based on: {{ critique1 }} Original: {{ v1 }} </step> <step as="critique2"> Critique v2: {{ v2 }} Any remaining issues? </step> <step> {{ if critique2 has issues }} Final revision based on: {{ critique2 }} {{ else }} {{ v2 }} {{ endif }} </step> \`\`\` --- ## Targeted Refinement \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="draft"> {{ task }} </step> <step as="clarity_check"> Rate clarity 1-10 and suggest fixes: {{ draft }} </step> <step as="accuracy_check"> Check for factual accuracy: {{ draft }} </step> <step as="tone_check"> Is the tone appropriate for {{ audience }}? {{ draft }} </step> <step> Final revision addressing: - Clarity: {{ clarity_check }} - Accuracy: {{ accuracy_check }} - Tone: {{ tone_check }} Original: {{ draft }} </step> \`\`\` --- ## When to Use | Content Type | Iterations | |--------------|------------| | Code | 1-2 | | Blog posts | 2-3 | | Legal/Medical | 3+ | | Simple responses | 0-1 | --- ## Best Practices - ✅ **DO**: Be specific in critique prompts - ✅ **DO**: Limit iterations (2-3 usually enough) - ✅ **DO**: Use different critique angles - ❌ **DON'T**: Iterate endlessly - ❌ **DON'T**: Use for simple tasks --- ## Next Steps - **Constitutional AI** → \`docs({ action: "get", topic: "technique-constitutional" })\` - **Chains** → \`docs({ action: "get", topic: "chains" })\` `; // ============================================================================ // TECHNIQUE-STEP-BACK // ============================================================================ export const DOCS_TECHNIQUE_STEP_BACK = `# Step-Back Prompting > Abstract first, then specific ## Overview Step-back prompting asks a high-level question first, then uses that understanding to answer the specific question. Improves reasoning on complex problems. --- ## Quick Reference \`\`\` Specific question → Step back to principles → Apply to specific \`\`\` --- ## Basic Step-Back \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="principles"> Before answering the specific question, let's step back. Question: {{ question }} What are the general principles or concepts needed to answer this? </step> <step> Using these principles: {{ principles }} Now answer the specific question: {{ question }} </step> \`\`\` --- ## Domain Step-Back \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="context"> Question: {{ question }} Step back: What domain knowledge is relevant here? - Key concepts: - Relevant rules: - Common patterns: </step> <step> Applying this context: {{ context }} Answer: {{ question }} </step> \`\`\` --- ## Example: Physics Problem \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <step as="physics_principles"> Problem: A ball is thrown upward at 20 m/s. How high does it go? Step back: What physics principles apply to projectile motion? </step> <step> Using: {{ physics_principles }} Solve: A ball is thrown upward at 20 m/s. How high does it go? </step> \`\`\` --- ## When to Use | Problem Type | Step-Back? | |--------------|------------| | Complex reasoning | Yes ✅ | | Domain-specific | Yes ✅ | | Multi-step logic | Yes ✅ | | Simple factual | No | | Creative tasks | No | --- ## Best Practices - ✅ **DO**: Use for problems requiring domain knowledge - ✅ **DO**: Keep step-back question focused - ✅ **DO**: Connect principles back to specific question - ❌ **DON'T**: Over-abstract simple questions - ❌ **DON'T**: Make step-back too generic --- ## Next Steps - **Chain of Thought** → \`docs({ action: "get", topic: "technique-cot" })\` - **Tree of Thoughts** → \`docs({ action: "get", topic: "technique-tot" })\` `; // ============================================================================ // TECHNIQUE-RAG // ============================================================================ export const DOCS_TECHNIQUE_RAG = `# RAG Technique > Retrieval-Augmented Generation fundamentals ## Overview RAG retrieves relevant context before generating responses. Grounds AI in real data, reducing hallucination and enabling knowledge-based answers. --- ## Quick Reference \`\`\` Query → Retrieve relevant docs → Generate with context \`\`\` --- ## RAG Components 1. **Retriever**: Finds relevant documents 2. **Context**: Retrieved information 3. **Generator**: Creates response using context --- ## Basic RAG Pattern \`\`\`promptl --- provider: OpenAI model: gpt-4o --- <system> Answer questions using ONLY the provided context. If the context doesn't contain the answer, say "I don't have information about that." </system> <user> Context: {{ context }} Question: {{ question }} </user> \`\`\` --- ## RAG with Tool \`\`\`promptl --- provider: OpenAI model: gpt-4o tools: - latitude/search --- <system> You are a research assistant. 1. Search for relevant information 2. Cite sources in your answer 3. Say "I couldn't find information" if search fails </system> <user> {{ question }} </user> \`\`\` --- ## RAG with Citation \`\`\`promptl --- provider: OpenAI model: gpt-4o schema: type: object properties: answer: type: string citations: type: array items: type: object properties: claim: type: string source: type: string required: [claim, source] required: [answer, citations] --- <system> Answer questions with citations. Every factual claim must have a citation. </system> <user> Context: {{ context }} Question: {{ question }} </user> \`\`\` --- ## Multi-Source RAG \`\`\`promptl --- provider: OpenAI model: gpt-4o tools: - search_docs: description: Search documentation parameters: type: object properties: query: type: string required: [query] - search_web: description: Search the web parameters: type: object properties: query: type: string required: [query] --- <system> Research assistant with multiple sources. 1. First check internal docs 2. Then search web if needed 3. Synthesize information from all sources </system> <user> {{ question }} </user> \`\`\` --- ## When to Use RAG | Scenario | RAG? | |----------|------| | Knowledge-based Q&A | Yes ✅ | | Current information | Yes ✅ | | Domain expertise | Yes ✅ | | Creative writing | No | | General conversation | No | --- ## Best Practices - ✅ **DO**: Instruct to use ONLY provided context - ✅ **DO**: Require citations - ✅ **DO**: Handle "no information" cases - ❌ **DON'T**: Trust model's pre-training for facts - ❌ **DON'T**: Skip context relevance check --- ## Next Steps - **RAG Recipe** → \`docs({ action: "get", topic: "recipe-rag" })\` - **Built-in tools** → \`docs({ action: "get", topic: "tools-builtin" })\` `;