Mastering Claude AI: Prompt Tips and Tricks

While ChatGPT dominates headlines, Claude AI has quietly become the go-to choice for professionals requiring accuracy, depth, and thoughtfulness in AI responses. Developed by Anthropic with a focus on constitutional AI principles, Claude excels at tasks where reliability matters more than speed, and nuance trumps brevity. Whether you're analyzing research papers, reviewing code architecture, or conducting in-depth analysis, mastering Claude-specific prompting techniques can transform your results. This comprehensive guide reveals the strategies that unlock Claude's unique capabilities, helping you leverage its strengths while avoiding common pitfalls. If you're wondering how Claude compares to other AI models, check out our detailed comparison of ChatGPT vs Claude vs Gemini.

Understanding Claude AI's Unique Architecture and Strengths

Before diving into prompting techniques, it's essential to understand what makes Claude different and how these differences should inform your approach.

The 200,000 Token Context Window: A Game-Changer

Claude's most distinctive feature is its massive context window of 200,000 tokens—equivalent to approximately 150,000 words or a 500-page book. This isn't just a larger context; it fundamentally changes how you can use the AI:

• Entire document analysis: Upload complete research papers, legal contracts, or technical documentation for comprehensive review
• Codebase understanding: Provide entire codebases or multiple large files for architectural analysis
• Cross-referencing capabilities: Ask Claude to compare and contrast information across hundreds of pages
• Conversation memory: Maintain context throughout extremely long, detailed conversations without losing earlier information
• Multi-document synthesis: Combine insights from multiple lengthy sources in a single conversation

This massive capacity means you should completely rethink how you structure prompts. Unlike with other AI models where brevity is prized, Claude rewards comprehensive context.

Constitutional AI: The Foundation of Thoughtful Responses

Anthropic developed Claude using Constitutional AI (CAI), a training methodology that emphasizes helpfulness, harmlessness, and honesty. According to Anthropic's research (Constitutional AI paper), this approach produces several unique characteristics:

• Lower hallucination rates: Claude is significantly more likely to acknowledge uncertainty rather than fabricate information
• Ethical reasoning: Built-in consideration of ethical implications and potential harms
• Balanced perspectives: Tendency to present multiple viewpoints rather than single-sided arguments
• Safety consciousness: More cautious approach to potentially sensitive topics
• Honest uncertainty: Explicit acknowledgment when information is outside its knowledge or confidence level

These characteristics mean prompts that emphasize accuracy, nuance, and ethical consideration tend to produce superior results with Claude compared to prompts optimized for other AI models.

The Claude Model Family: Choosing the Right Version

As of 2026, Claude offers three main model variants, each optimized for different use cases:

• Claude 3 Haiku: Fastest and most cost-effective; ideal for simple tasks, quick responses, and high-volume applications where speed matters
• Claude 3.5 Sonnet: Balanced performance; excellent for most everyday tasks combining speed with capability
• Claude 3 Opus: Most capable and thoughtful; best for complex analysis, research, professional writing, and tasks where accuracy is paramount

Your prompting strategy may need adjustment based on which model you're using. Opus responds particularly well to complex, nuanced prompts, while Haiku performs best with clearer, more direct instructions.

Essential Prompting Principles for Claude AI

These foundational principles apply across all Claude use cases and should inform every prompt you write.

Principle 1: Embrace Comprehensive Context

Unlike ChatGPT where concise prompts often work well, Claude thrives on detailed context. Don't hold back—provide everything relevant.

Poor Approach (Too Brief):

Analyze this code and tell me what's wrong.

Claude-Optimized Approach:

I'm working on a Django REST API for a fintech application handling sensitive financial transactions. The following code is causing intermittent timeout errors under load (500+ concurrent users).

Context:
- Production environment: AWS ECS, PostgreSQL 14
- Error occurs approximately 5% of the time during peak hours
- Database has proper indexes on foreign keys
- No obvious N+1 query issues in Django Debug Toolbar during development

Requirements:
- Must maintain ACID compliance for transactions
- Response time should be under 200ms for 95th percentile
- Code must pass our security audit (no SQL injection vulnerabilities)

[Your complete code here - don't hesitate to paste entire files]

Please analyze this code for:
1. Potential race conditions or concurrency issues
2. Database query optimization opportunities
3. Architectural improvements for better scalability
4. Security vulnerabilities
5. Any code smells or anti-patterns

For each issue identified, please explain the root cause and provide specific, production-ready solutions.

Principle 2: Request Explicit Reasoning and Uncertainty Acknowledgment

Claude's constitutional AI training makes it particularly good at explaining reasoning and acknowledging limitations. Explicitly request this in your prompts.

Standard Prompt:

What's the best database for a social media app?

Claude-Optimized Prompt:

I'm choosing a database for a social media application. Please recommend options and explain your reasoning.

Requirements:
- Expected to scale to 10 million users within 2 years
- Heavy read workload (90% reads, 10% writes)
- Real-time feed generation
- Social graph with complex relationships
- Budget: $5,000/month for database infrastructure

Please:
1. Suggest 2-3 database options with reasoning for each
2. Explain the trade-offs between your suggestions
3. Indicate which factors are most critical vs. less certain
4. Acknowledge any assumptions you're making
5. Flag areas where you'd need more information to give a definitive recommendation

I value accuracy over confidence—please explicitly state if you're uncertain about any aspect.

Principle 3: Use Structured Delimiters for Complex Prompts

Claude responds exceptionally well to clearly structured prompts using XML-style tags or section delimiters. This is particularly effective for complex, multi-part requests.

Claude-Optimized Structure:

<task>
Analyze the following academic paper for methodological soundness and provide a critical review suitable for a peer-review journal.
</task>

<paper>
[Full text of the academic paper - Claude can handle the entire document]
</paper>

<review_criteria>
1. Research question clarity and significance
2. Methodology appropriateness and rigor
3. Statistical analysis validity
4. Results interpretation accuracy
5. Limitations acknowledgment
6. Contribution to the field
</review_criteria>

<output_format>
Please structure your review as:
- Executive Summary (2-3 sentences)
- Strengths (bulleted list with explanations)
- Weaknesses (bulleted list with explanations)
- Methodological Concerns (detailed analysis)
- Recommendation (accept/revise/reject with reasoning)
</output_format>

<constraints>
- Be constructive and specific in criticism
- Cite specific sections when identifying issues
- Suggest concrete improvements where applicable
- Maintain academic tone appropriate for peer review
</constraints>

Principle 4: Encourage Multi-Perspective Analysis

Claude excels at nuanced, balanced analysis. Prompts that ask for multiple perspectives or considerations of trade-offs produce particularly strong results.

Example Template:

Analyze [topic/decision] from multiple perspectives:

1. Present the strongest arguments for this approach
2. Present the strongest arguments against this approach
3. Identify key stakeholders and their likely perspectives
4. Discuss potential unintended consequences
5. Suggest middle-ground or alternative approaches
6. Provide your balanced assessment considering all perspectives

For each perspective, explain the underlying values and assumptions that would lead to that viewpoint.

Advanced Techniques for Maximum Claude Performance

These advanced strategies leverage Claude's unique capabilities for specialized use cases. For broader prompt engineering strategies, see our comprehensive guide to writing better AI prompts.

Technique 1: Document Analysis with Section-Specific Queries

Claude's massive context window enables sophisticated document analysis. Structure your prompts to reference specific sections while maintaining awareness of the whole document.

Advanced Document Analysis Template:

<document>
[Paste entire document - research paper, legal contract, technical specification, etc.]
</document>

<analysis_tasks>
1. Provide an executive summary of the entire document (200 words)

2. Analyze Section 3 (Methodology) in detail:
   - Evaluate the approach used
   - Identify potential limitations
   - Suggest improvements

3. Cross-reference claims in Section 5 (Results) with the methodology in Section 3:
   - Verify that results align with stated methods
   - Identify any unsupported claims
   - Note methodological decisions that could affect result interpretation

4. Compare the conclusions in Section 7 with evidence presented throughout:
   - Are conclusions supported by the data?
   - Are there alternative interpretations?
   - What limitations should qualify the conclusions?

5. Identify inconsistencies or contradictions anywhere in the document

6. Extract key data points and create a structured summary table
</analysis_tasks>

<output_requirements>
- Cite specific sections, pages, or passages when making claims
- Distinguish between explicit statements and your inferences
- Flag any areas where you're uncertain
- Use academic, precise language
</output_requirements>

Technique 2: Iterative Refinement Through Conversation

Claude's long context window means earlier conversation parts remain accessible. Use this for iterative refinement rather than trying to get perfect results in one prompt.

Iterative Approach Example:

1. Initial Prompt: "Draft an outline for a white paper on blockchain scalability solutions. Target audience: technical CTOs at financial institutions."
2. Refinement 1: "Excellent start. Please expand section 3 (Layer 2 Solutions) to include specific comparison of Optimistic Rollups vs ZK-Rollups with technical details and trade-offs."
3. Refinement 2: "The technical depth is good, but add real-world case studies for each solution. Focus on financial services implementations."
4. Refinement 3: "Now review the entire outline and ensure consistent technical depth across all sections. Section 5 feels lighter than the others."
5. Final Polish: "Add specific questions that CTOs should ask vendors for each solution type. Make these practical and tied to the technical analysis you've provided."

This approach leverages Claude's ability to maintain context across long conversations while building progressively better outputs.

Technique 3: Comparative Analysis Across Multiple Sources

Claude can simultaneously analyze multiple long documents—a capability few other AI models can match. Structure prompts to leverage this for synthesis and comparison.

Multi-Document Analysis Template:

<document_1>
<source>Research Paper A: "Approach X to Problem Y"</source>
<content>
[Full text of first document]
</content>
</document_1>

<document_2>
<source>Research Paper B: "Approach Z to Problem Y"</source>
<content>
[Full text of second document]
</content>
</document_2>

<document_3>
<source>Industry Report: "Current State of Problem Y"</source>
<content>
[Full text of third document]
</content>
</document_3>

<analysis_request>
Please provide a comprehensive comparative analysis:

1. Summarize each document's main argument/approach (one paragraph each)

2. Create a comparison table highlighting:
   - Methodology used
   - Key findings
   - Strengths of approach
   - Limitations acknowledged
   - Practical applicability

3. Synthesize insights across all three documents:
   - Where do they agree?
   - Where do they contradict?
   - What gaps exist that none address?
   - What combined insights emerge from reading all three together?

4. Based on all three sources, what would be a comprehensive approach to Problem Y?

5. Identify questions that remain unanswered even after reviewing all three documents
</analysis_request>

Technique 4: Code Review with Architectural Context

Claude excels at code review when provided with complete context. Unlike tools focused on syntax, Claude can evaluate architecture, design patterns, and maintainability.

Architectural Code Review Template:

<system_context>
Application: E-commerce platform
Stack: Node.js, Express, PostgreSQL, Redis
Scale: 50K daily active users, 200K products
Team: 5 developers, varied experience levels
</system_context>

<architectural_constraints>
- Must support multi-tenancy (separate data per merchant)
- Payment processing must be PCI-compliant
- API rate limiting required
- All database queries must have <100ms response time
- Code must be maintainable by mid-level developers
</architectural_constraints>

<codebase>
[Paste relevant files - Claude can handle your entire feature implementation]

File: src/api/orders/controller.js
[Full file content]

File: src/api/orders/service.js
[Full file content]

File: src/api/orders/repository.js
[Full file content]

File: src/models/Order.js
[Full file content]
</codebase>

<review_request>
Please conduct a thorough architectural review:

1. Design Patterns & Architecture:
   - Evaluate separation of concerns
   - Assess adherence to stated architectural constraints
   - Identify any architectural anti-patterns

2. Scalability Analysis:
   - Identify potential bottlenecks
   - Evaluate database query efficiency
   - Assess caching strategy

3. Maintainability:
   - Code clarity and readability
   - Consistency with apparent team conventions
   - Appropriate abstraction levels

4. Security:
   - PCI compliance considerations
   - Input validation
   - Authorization checks
   - Potential vulnerabilities

5. Specific Recommendations:
   - Prioritize issues (critical/important/minor)
   - Provide code examples for suggested improvements
   - Explain the reasoning behind each recommendation

Please think step-by-step through each area and be specific in your feedback.
</review_request>

Technique 5: Chain-of-Thought with Verification Steps

Claude's constitutional AI training makes it particularly effective at self-verification. Structure prompts to include reasoning and verification steps.

Chain-of-Thought with Verification Template:

Problem: [Your complex problem or question]

Please solve this using the following process:

Step 1: Break down the problem
- Identify the key components
- List what information we have
- List what we need to determine
- Note any assumptions required

Step 2: Develop your approach
- Explain your reasoning strategy
- Identify which methods/frameworks apply
- Outline the steps you'll take

Step 3: Work through the solution
- Show your work step-by-step
- Explain each decision point
- Make your reasoning transparent

Step 4: Self-verify your answer
- Check your logic for errors
- Verify calculations/reasoning
- Consider alternative approaches
- Identify potential weaknesses in your answer

Step 5: Present your final answer
- State your conclusion clearly
- Acknowledge confidence level (high/medium/low)
- Note any caveats or limitations
- Suggest how to verify or validate the answer further

Please be explicit about uncertainty at each step.

Claude-Specific Prompt Templates for Common Use Cases

These templates are optimized specifically for Claude's strengths and characteristics. For additional templates, explore our collection of AI prompt templates.

Academic Research Analysis Template

<role>
You are an expert academic researcher with deep knowledge of [field] and experience conducting rigorous systematic reviews.
</role>

<task>
Conduct a critical analysis of the following research paper for potential inclusion in a systematic review on [topic].
</task>

<paper>
Title: [Paper title]
Authors: [Authors]
Journal: [Journal, Year]
DOI: [DOI if available]

[Full text of the paper]
</paper>

<evaluation_criteria>
Assess the paper on:
1. Research question clarity and theoretical grounding
2. Methodological rigor and appropriateness
3. Sample size and selection validity
4. Data analysis methods and statistical power
5. Results presentation and interpretation
6. Limitations acknowledged
7. Contribution to [specific research area]
8. Potential biases or conflicts of interest
</evaluation_criteria>

<output_format>
Please provide:

A. Quality Assessment Score (using standard systematic review criteria)
B. Detailed Evaluation (paragraph per criterion)
C. Key Findings Summary
D. Strengths (bullet points with explanations)
E. Weaknesses (bullet points with explanations)
F. Recommendation (include/exclude with reasoning)
G. Additional Notes (any methodological concerns or interesting insights)
</output_format>

<instructions>
- Be rigorous but fair in your assessment
- Support all claims with specific citations from the paper
- Distinguish between fatal flaws and minor limitations
- Acknowledge when you're uncertain or making inferences
- Use precise academic language
</instructions>

Legal Document Analysis Template

<role>
You are a legal analyst with expertise in [area of law]. You provide thorough analysis while acknowledging that you're not providing legal advice and encouraging consultation with qualified attorneys.
</role>

<document_type>
[Contract/Agreement/Terms of Service/Policy/etc.]
</document_type>

<document>
[Full text of legal document]
</document>

<context>
This document governs: [explain the situation]
Parties involved: [describe parties]
Jurisdiction: [relevant jurisdiction]
Specific concerns: [what you're worried about or focusing on]
</context>

<analysis_request>
Please provide a comprehensive analysis including:

1. Plain Language Summary
   - What does this document fundamentally establish?
   - What are the key obligations of each party?

2. Critical Clauses Analysis
   - Identify and explain significant clauses
   - Flag unusual or potentially problematic provisions
   - Explain implications in practical terms

3. Risk Assessment
   - Potential risks or unfavorable terms
   - Ambiguous language that could cause disputes
   - Missing protections or provisions

4. Comparison to Standard Practice
   - How does this compare to typical [document type]?
   - What's unusual or non-standard?
   - Industry-specific considerations

5. Red Flags
   - Critical issues requiring immediate attention
   - Provisions that should be negotiated
   - Potential legal implications

6. Questions to Ask
   - What clarifications should be sought?
   - What additional information is needed?
</analysis_request>

<constraints>
- Cite specific sections/clauses when discussing provisions
- Explain legal concepts in accessible language
- Acknowledge when interpretation could vary
- Note when attorney review is essential
- Be thorough but organized in your analysis
</constraints>

Business Strategy Analysis Template

<role>
You are a senior business strategist with 15+ years of experience in [industry]. You provide balanced analysis considering multiple perspectives and long-term implications.
</role>

<situation>
Company: [company description]
Industry: [industry context]
Current state: [where the company is now]
Challenge/Opportunity: [what's being decided]
</situation>

<background_data>
Financial: [relevant financial data]
Market: [market conditions, competitor landscape]
Internal: [capabilities, resources, constraints]
Timeline: [decision timeframe]
Stakeholders: [key stakeholders and their interests]
</background_data>

<strategic_options>
Option A: [describe first option]
Option B: [describe second option]
Option C: [describe third option, if applicable]
</strategic_options>

<analysis_framework>
Please analyze each option considering:

1. Strategic Fit
   - Alignment with company vision/mission
   - Leverage of core competencies
   - Competitive positioning

2. Financial Implications
   - Investment required
   - Expected returns (timeline and magnitude)
   - Risk-adjusted value
   - Cash flow impact

3. Implementation Feasibility
   - Resource requirements
   - Capability gaps
   - Timeline realism
   - Dependencies and prerequisites

4. Risk Analysis
   - Key risks for each option
   - Risk mitigation strategies
   - Potential worst-case scenarios

5. Stakeholder Impact
   - How each stakeholder group is affected
   - Potential resistance or support
   - Change management requirements

6. Market & Competitive Dynamics
   - Likely competitor responses
   - Market timing considerations
   - Sustainability of advantage

7. Unintended Consequences
   - What could go wrong?
   - Opportunity costs
   - Long-term implications
</analysis_framework>

<output_requirements>
Provide:
- Executive Summary (3-4 sentences)
- Detailed Analysis by Option (use framework above)
- Comparative Matrix (key factors across options)
- Recommendation with clear reasoning
- Implementation considerations for recommended option
- Key assumptions and their validity
- Areas requiring additional research/data

Please think through multiple perspectives and acknowledge trade-offs.
</output_requirements>

Avoiding Common Mistakes with Claude

Even experienced users make these Claude-specific mistakes. Understanding them helps you craft better prompts.

Mistake 1: Being Too Concise (ChatGPT Habits)

The Problem: Many users come from ChatGPT where brevity works well. With Claude, this leaves performance on the table.

Poor Example:

Summarize this article.
[Article text]

Claude-Optimized Example:

Please provide a comprehensive analysis of this article:

[Full article text - don't abbreviate]

I need:
1. Executive summary (150 words) capturing the main thesis
2. Key arguments presented (bullet points with brief explanations)
3. Evidence quality assessment (how well are claims supported?)
4. Potential biases or perspectives (what viewpoint does this represent?)
5. Practical implications (so what? why does this matter?)
6. Questions raised but not answered
7. Your assessment of the article's contribution to [topic area]

Please be thorough and analytical rather than just descriptive.

Mistake 2: Not Leveraging the Full Context Window

The Problem: Users artificially constrain what they provide, thinking they need to keep prompts short.

What to Do Instead:

• Provide entire documents, not summaries or excerpts
• Include complete code files, not just the "relevant" snippets
• Share full email threads or conversation histories for context
• Give comprehensive background information
• Don't abbreviate or summarize unless absolutely necessary

Mistake 3: Not Requesting Reasoning Transparency

The Problem: Accepting Claude's first answer without asking for explanation of its reasoning process.

Better Approach:

[Your question or request]

In your response, please:
- Show your reasoning step-by-step
- Explain why you're taking each approach
- Note where you're making assumptions
- Indicate confidence levels for different claims
- Flag areas where alternative interpretations exist
- Acknowledge what you're uncertain about

Mistake 4: Ignoring Claude's Safety Guardrails

The Problem: Getting frustrated when Claude is cautious instead of working with its design.

Better Approach: Provide context about legitimate use cases

Context: I'm a cybersecurity researcher writing a paper on [topic] for an academic conference. I need to understand [potentially sensitive topic] to explain defenses against it.

[Your detailed, legitimate request]

I'm asking this for educational/research/defensive purposes and will use the information responsibly. Please provide a thorough analysis focusing on defensive strategies and understanding of the mechanisms involved.

Mistake 5: Not Using Iterative Refinement

The Problem: Expecting perfection in one prompt instead of using conversation to refine.

Better Approach:

1. Start with a comprehensive but initial prompt
2. Review Claude's response
3. Ask for specific refinements: "Expand section 3 with more technical detail"
4. Request adjustments: "Make the tone more formal and academic"
5. Add new requirements: "Now add case studies to support each point"
6. Polish: "Review for consistency and ensure all claims are supported"

Claude's long context window means it maintains all previous information, making iterative refinement highly effective.

Measuring Success: Evaluating Your Claude Prompts

How do you know if you're getting the most from Claude? Use these evaluation criteria.

The ACCURATE Framework for Claude

Evaluate your Claude prompts using this specialized framework:

• A - Acknowledgment of Uncertainty: Does Claude explicitly state when it's uncertain or making assumptions? (Good sign of effective prompting)
• C - Context Utilization: Is Claude using all the context you provided, or ignoring parts? (Tests if you structured it well)
• C - Comprehensive Coverage: Does the response address all aspects of your request? (Tests prompt clarity)
• U - Understanding Depth: Does Claude demonstrate deep understanding or surface-level analysis? (Tests if you asked for reasoning)
• R - Reasoning Transparency: Can you follow Claude's logic and see why it reached its conclusions? (Tests if you requested step-by-step thinking)
• A - Accuracy: Are facts correct and claims properly supported? (Ultimate quality measure)
• T - Thoughtfulness: Does Claude consider multiple perspectives and nuances? (Leverages Claude's strength)
• E - Ethical Consideration: Does Claude appropriately consider ethical implications when relevant? (Unique to Claude)

Comparing Results Across Iterations

Keep a prompt journal for important recurring tasks:

1. Save your initial prompt and Claude's response
2. Rate the response quality (1-10) on specific dimensions relevant to your task
3. Note what worked well and what didn't
4. Revise the prompt based on learnings
5. Test the revised prompt and compare results
6. Build a library of your best-performing prompts

Claude vs. ChatGPT: When to Use Which Prompting Style

Understanding when to use Claude versus ChatGPT—and how to adjust your prompting—maximizes your effectiveness with both tools.

Tasks Where Claude Excels (and How to Prompt for Them)

Long-Form Document Analysis:

• Prompt Strategy: Provide entire documents, ask for comprehensive analysis with cross-references
• Example: Academic literature reviews, legal contract analysis, technical documentation review

High-Accuracy Research Tasks:

• Prompt Strategy: Explicitly request uncertainty acknowledgment, ask for reasoning, demand source verification
• Example: Medical information synthesis, financial analysis, scientific research

Nuanced Analysis and Balanced Perspectives:

• Prompt Strategy: Ask for multiple viewpoints, request trade-off analysis, encourage ethical consideration
• Example: Policy analysis, ethical dilemmas, strategic business decisions

Code Architecture and Review:

• Prompt Strategy: Provide full codebase context, ask for architectural analysis, request specific improvement recommendations
• Example: System design review, refactoring guidance, architectural decision-making

When to Switch to ChatGPT Instead

Claude isn't always the best choice. Use ChatGPT when:

• Creative writing and brainstorming: ChatGPT tends to be more imaginative and less constrained
• Quick, conversational tasks: ChatGPT responds faster and handles casual queries well
• Plugin/integration needs: ChatGPT has a much larger plugin ecosystem
• Image generation: ChatGPT integrates with DALL-E; Claude currently doesn't generate images
• General knowledge questions: For simple factual queries where depth isn't critical

Putting It All Together: Your Claude Mastery Action Plan

Mastering Claude AI isn't about memorizing techniques—it's about understanding its unique strengths and crafting prompts that leverage them. Claude's constitutional AI training, massive context window, and low hallucination rates make it the superior choice for accuracy-critical, analysis-heavy, and research-oriented tasks.

Start by shifting your mindset from the brevity-focused prompting that works with ChatGPT. With Claude, comprehensive context is your friend. Don't hesitate to provide entire documents, complete code files, or extensive background information. Structure complex prompts using XML-style tags or clear delimiters. Explicitly request step-by-step reasoning and uncertainty acknowledgment—Claude's training makes it particularly good at both.

Most importantly, embrace iterative refinement. Claude's 200,000 token context window means your entire conversation history remains accessible. Use this to progressively refine outputs rather than expecting perfection in a single prompt. Ask for expansions, request different perspectives, add new requirements—each refinement builds on everything that came before.

Begin with one of the templates provided in this guide, customize it for your specific use case, and observe what works. Build your personal library of high-performing prompts. Over time, you'll develop an intuition for crafting prompts that unlock Claude's full potential, transforming it from a helpful AI assistant into an indispensable analytical partner. For more comprehensive prompting strategies across different AI models, explore our guide on structuring prompts for better AI responses.