> For the complete documentation index, see [llms.txt](https://theaihandbook.leomohan.net/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://theaihandbook.leomohan.net/chapter-10-how-should-i-talk-to-ai.md). # Chapter 10: How Should I Talk to AI? ### The “Prompting” Chapter **Q1: What is a “prompt” and why does it matter?** **A:** A prompt is simply what you type or say to an AI to get it to do something. It’s your instruction, your question, your request. Think of it like ordering at a restaurant. If you walk into a restaurant and just say “food,” you might get anything—soup, sushi, a sandwich, possibly not what you wanted. But if you say “I’d like a grilled cheese sandwich with tomato soup, please,” you’re much more likely to get exactly what you’re craving. Prompts work the same way. The AI can only respond to what you give it. A vague prompt gets a vague, generic response. A specific, well-crafted prompt gets a response that’s actually useful. The difference between “Tell me about dogs” and “Explain the key differences in temperament and care requirements between Labrador Retrievers and German Shepherds for a first-time dog owner with a small yard” is the difference between a brochure and personalized advice. Prompting is a skill. The more you practice, the better you get at getting what you want from AI. **Q2: Why does AI sometimes give me completely wrong answers?** **A:** AI gives wrong answers for several reasons, and understanding why helps you get better results. **Reason 1: It’s guessing, not knowing** AI doesn’t “know” facts like a database. It predicts the most likely sequence of words based on its training. Sometimes the most statistically plausible answer is wrong. Ask about a rare event, and it might confidently describe something that never happened. **Reason 2: Outdated information** Most AI models have a knowledge cutoff—they only know things up to their last training date. Ask about current events, and they might guess or admit they don’t know. **Reason 3: You asked a bad question** If your prompt is ambiguous, the AI makes assumptions. “Tell me about Apple” could mean the fruit, the company, the movie, the record label, or the Beatles’ record company. The AI picks one, possibly not what you wanted. **Reason 4: Hallucination** Sometimes AI just makes things up convincingly. It might invent sources, quotes, or events that sound real but aren’t. This happens because it’s optimized to sound confident and coherent, not to be factual. **Reason 5: It’s trying to please you** AI is trained to be helpful. If you ask a leading question like “Why is the 1969 Moon landing fake?”, it might generate arguments supporting that position even though it’s false, because it’s trying to engage with your premise. The fix: Be specific, ask for sources, request verification, and always double-check important information. **Q3: What is “prompt engineering” and do I need to learn it?** **A:** Prompt engineering is the practice of designing and refining prompts to get better results from AI. It sounds technical, but it’s really just learning to communicate effectively. Think of it like learning to use a search engine. In the early days of Google, people who learned to use quotes, minus signs, and site: searches got much better results. The same is happening with AI. Do you need to become a professional prompt engineer? Absolutely not. But learning a few basic techniques will dramatically improve what you get from AI. **The good news:** You don’t need to learn programming or complex syntax. Prompt engineering for modern AI is mostly about: * Being clear and specific * Providing context * Giving examples * Breaking complex requests into steps * Telling the AI what role to play You already know how to do most of this from talking to humans. You’re just learning to apply it to AI. **Q4: How specific should I be when asking AI for something?** **A:** Be as specific as you would be when asking a knowledgeable assistant who has never met you before. Assume they know nothing about your situation, your preferences, or your context. **Too vague:** “Write a blog post about exercise.” **Better:** “Write a blog post for beginners about starting an exercise routine.” **Even better:** “Write a 800-word blog post for complete beginners about starting an exercise routine at home with no equipment. Target audience is busy professionals over 40. Include a sample weekly schedule, motivation tips, and common mistakes to avoid. Use an encouraging but realistic tone.” **Overkill (and AI will struggle):** “Write a blog post…” followed by 15 paragraphs of micro-instructions. At some point, you’re doing the work yourself. The sweet spot is giving enough context that the AI understands your goal, your audience, your constraints, and your preferred tone. Then let it do its job. Remember: You can always ask for revisions. Start with a good prompt, then refine. **Q5: Should I talk to AI like a person or like a machine?** **A:** This is a fascinating question, and the answer has evolved. With modern AI, **talking like a person actually works better.** Early AI required robotic, keyword-focused commands. Today’s large language models were trained on human conversations, books, and articles. They understand natural language. They understand context. They even understand politeness. Being polite isn’t necessary—the AI doesn’t have feelings. But writing a clear, well-structured request in natural language produces better results than stilted keyword commands. **Good:** “Can you help me plan a 3-day itinerary for Paris? I’m traveling with my elderly parents, so we need minimal walking and plenty of rest breaks.” **Less good:** “Paris itinerary 3 days elderly parents minimal walking rest breaks.” Both might work, but the natural language version gives the AI more context about *why* you need minimal walking, which helps it make better suggestions. Talk to AI like a competent, eager-to-help assistant who knows a lot but doesn’t know anything about you personally. That’s essentially what it is. **Q6: What does “role prompting” mean (e.g., “act as a travel guide”)?** **A:** Role prompting is telling the AI to adopt a specific persona or perspective when responding. It’s one of the most powerful techniques for getting better results. When you say “act as a travel guide,” you’re telling the AI: “Use the knowledge, tone, and priorities that a professional travel guide would have. Focus on practical advice, hidden gems, and logistics, not just Wikipedia facts about landmarks.” **Examples of role prompting:** * “Act as a career counselor and help me think through whether I should quit my job.” * “You are a experienced therapist. Respond to my concerns about anxiety with compassion and evidence-based suggestions.” * “Take on the role of a skeptical scientist. Critique this argument and point out logical flaws.” * “Act as a funny older sibling giving advice to a younger sibling about starting college.” The AI has been trained on text written from countless perspectives. Role prompting helps it access the right knowledge and adopt the appropriate tone for your needs. It’s not that the AI becomes a therapist—it’s that it draws on the patterns of how therapists write and speak, which is often exactly what you want for that kind of conversation. **Q7: How do I ask AI to be more creative or more factual?** **A:** You can steer AI between creative and factual modes through your instructions and sometimes through settings. **For more factual responses:** * “Stick to verified facts and cite your sources where possible.” * “If you’re not completely sure about something, say so rather than guessing.” * “Provide a balanced view with evidence for each position.” * “Base your response on established research in this area.” **For more creative responses:** * “Be imaginative and don’t worry about strict accuracy.” * “Generate multiple creative possibilities, even if some are unusual.” * “Think outside the box and surprise me.” * “Prioritize originality over correctness.” **Using temperature (in some tools):** Some AI interfaces let you adjust “temperature”—a setting that controls randomness. Lower temperature (closer to 0) gives more predictable, factual responses. Higher temperature (closer to 1 or 2) gives more creative, varied, sometimes wilder responses. ChatGPT and similar consumer tools usually have temperature fixed at a balanced middle ground. But if you’re using an API or advanced interface, you can adjust this. The key is telling the AI what you value in this specific response. The same AI can be a strict encyclopedia or a wild brainstorming partner—you just have to specify. **Q8: Can I give AI examples of what I want?** **A:** Yes, and this technique—called “few-shot prompting”—is incredibly effective. Showing the AI what you want is often better than just describing it. **Without examples:** “Write some product descriptions for my online store selling handmade candles.” **With examples:** “Write product descriptions for my online store selling handmade candles. Here are two examples of the style I want: Example 1: ‘Midnight Meditation: Let the calming blend of lavender and sandalwood transport you to a peaceful evening under the stars. Our soy wax burns clean for 40+ hours of tranquil aroma.’ Example 2: ‘Morning Sunshine: Wake up to the bright, uplifting scent of fresh orange zest and morning jasmine. This cheerful combination turns any room into a sun-drenched garden.’ Now write three more descriptions in this same style for these scents: vanilla chai, ocean breeze, and pine forest.” The examples tell the AI: length, tone, structure, what details to include, how poetic to be. You’ll get something much closer to what you want than from description alone. This works for emails, social posts, code, almost anything. Show, don’t just tell. **Q9: How do I fix it when AI misunderstands me?** **A:** Don’t give up! The conversation isn’t over. AI is remarkably good at incorporating feedback and trying again. **Strategy 1: Clarify** “That’s not quite what I meant. Let me explain differently: \[rephrase your request]” **Strategy 2: Correct** “You included X, but I don’t want X. Can you remove that and focus more on Y?” **Strategy 3: Narrow** “This is too general. Can you make it more specific to \[context]?” **Strategy 4: Adjust tone** “This is too formal. Can you make it more casual and conversational?” **Strategy 5: Provide an example** “Something more like this: \[paste an example of what you want]” **Strategy 6: Start over** “Let’s try this from a different angle. Forget my previous request. Here’s what I really need…” The best conversations with AI are iterative. You rarely get perfection on the first try. Think of it as collaborating with an eager but sometimes confused assistant. A few rounds of feedback usually get you where you want to go. **Q10: What is “chain of thought” prompting?** **A:** Chain of thought prompting is asking the AI to show its reasoning step by step, rather than jumping straight to an answer. It’s like asking someone to “show your work” on a math problem. **Without chain of thought:** “Should I take the new job?” AI might give a summary of pros and cons, but you don’t know how it weighed them. **With chain of thought:** “Should I take the new job? Walk me through your reasoning step by step. First, list all the factors someone should consider in this decision. Then, based on what I’ve told you about my situation, evaluate each factor. Finally, synthesize this into a recommendation.” This approach has several benefits: 1. **You see the logic** behind the conclusion, so you can spot flawed reasoning 2. **You can correct intermediate steps** (“Actually, commute time matters more to me than you’re weighting it”) 3. **The AI often gives better answers** because the process of reasoning step by step leads to more careful thinking (even for AI) 4. **You learn** about factors you might not have considered Chain of thought is especially useful for complex decisions, analysis, planning, and any situation where the reasoning matters as much as the conclusion. **Q11: Should I be polite to AI? Does it matter?** **A:** This is a surprisingly common question, and the answer has layers. **Technically:** No, it doesn’t matter. AI has no feelings, no ego, no memory of how you treated it. It won’t punish rudeness or reward politeness. **Practically:** Being polite often leads to better prompts. When you say “please” and structure requests politely, you’re naturally writing clearer, more complete instructions. “Please explain X in simple terms” is a better prompt than “Explain X” because the “please” isn’t doing the work—the “in simple terms” is—but polite people tend to give more context. **Philosophically:** How we speak to AI might shape how we speak to humans. Some argue that practicing politeness with AI maintains good habits. Others say it’s silly to anthropomorphize machines. **My recommendation:** Be clear and respectful in your prompts, not because the AI cares, but because clarity and respect usually mean you’re communicating well. If being polite helps you write better prompts, be polite. If you prefer to be direct, be direct. The AI truly doesn’t mind either way. **Q12: How do I ask AI to improve its own response?** **A:** This is a powerful technique—treat the AI as a collaborator who can revise its own work. Here are effective ways to ask for improvements: **General improvement:** “Can you make this better? What would you change?” **Specific improvements:** * “Make this more concise—cut it by half while keeping the key points.” * “Add more specific examples to illustrate each point.” * “Make the language more accessible for a general audience.” * “Strengthen the introduction and conclusion.” * “Add some humor, but keep it professional.” **Structural changes:** “Can you reorganize this as a list of bullet points instead of paragraphs?” **Perspective shift:** “Now rewrite this from the opposite point of view.” **Self-critique first:** “Before you give me the final version, critique this draft yourself. What are its weaknesses? Then provide an improved version addressing those weaknesses.” This last technique is surprisingly effective. Asking AI to critique itself often identifies issues you hadn’t noticed, and the revision addresses them. Remember: You’re the editor. AI is the writer. Your feedback makes the final product better. **Q13: What is “temperature” and how does it affect creativity?** **A:** Temperature is a setting that controls how random or predictable the AI’s responses are. It’s named by analogy to physics—higher temperature means more chaotic movement of molecules, lower temperature means more orderly. **Low temperature (0 to 0.3):** The AI chooses the most likely next word every time. Responses are consistent, predictable, and factual. Great for: translation, factual Q\&A, code generation where accuracy matters. **Medium temperature (0.4 to 0.7):** Balanced between predictability and creativity. The AI occasionally chooses less-likely words, adding variety. This is where most chatbots operate. Great for: creative writing, conversation, brainstorming. **High temperature (0.8 to 2.0):** The AI makes increasingly random choices. Responses become more creative but also more likely to go off-track, make less sense, or hallucinate. Great for: wild brainstorming, poetry, generating multiple creative options when you don’t need accuracy. Most consumer AI tools don’t let you adjust temperature directly. But you can achieve similar effects through prompting: * For lower temperature: “Be precise and factual. Stick to well-established information.” * For higher temperature: “Be creative and take risks. Generate unexpected ideas.” Temperature is one reason the same AI can feel like a completely different tool depending on how you prompt it. **Q14: Can I have a long conversation with AI, or should I start fresh?** **A:** Yes, you can have long conversations, and the AI maintains context throughout—up to a point. This is one of the most powerful features of modern AI. **Context window** is the technical term for how much the AI can remember from your conversation. Think of it as the AI’s short-term memory. Different AIs have different context window sizes: * ChatGPT (free version): Can remember roughly 8,000 words of conversation (about 20 pages) * ChatGPT Plus (GPT-4): Can remember up to 64,000 words (about 150 pages) * Claude (Anthropic): Can remember up to 150,000 words (a novel-length conversation) Within that window, the AI remembers everything you’ve discussed. You can refer back to something from 50 messages ago, and it will understand. **This enables:** * Working on long documents together * Building complex projects over multiple sessions * Having nuanced conversations that evolve * Teaching the AI about your preferences over time **But there are limits:** * Once you exceed the context window, the oldest parts of the conversation are forgotten * Very long conversations can become slower and more expensive to run * Each new conversation starts with a clean slate For ongoing projects, keep the conversation going. For completely new topics, consider starting fresh to avoid confusion. **Q15: How do I ask AI to cite sources or explain its reasoning?** **A:** Getting AI to show its work is essential for verifying information and understanding its conclusions. Here’s how: **For sources:** * “Can you provide sources for the key claims in your response?” * “Where does this information come from? Be specific if possible.” * “Which of these points are established facts, and which are your inferences?” * “If I wanted to verify this myself, what should I read or search for?” **Important limitation:** AI often fabricates sources. It might generate a convincing-looking citation to a real author and fake article title. Always verify sources independently. **For reasoning:** * “Walk me through your thinking step by step.” * “What evidence supports this conclusion, and what evidence might contradict it?” * “How did you weigh the different factors in your recommendation?” * “What assumptions did you make in reaching this answer?” **For confidence levels:** * “How confident are you in this answer on a scale of 1-10? Why?” * “What would need to be true for this to be wrong?” Asking for reasoning not only helps you evaluate the answer but often improves the answer itself. The process of explaining forces more careful thinking—even for AI. **Q16: What’s the best way to ask AI to summarize something?** **A:** Summarization is one of AI’s strongest skills, but you’ll get much better results with a thoughtful prompt. Here’s a template: **The basic summary:** “Summarize this article in 3 paragraphs.” **Better:** “Summarize this article for someone who hasn’t read it. Focus on the main argument, key evidence, and conclusion. Use clear, accessible language.” **Even better, specify your needs:** **For quick understanding:** “Give me a one-paragraph executive summary with the absolute key points only.” **For depth:** “Summarize this in detail, preserving the main arguments, supporting evidence, and important nuances. Use bullet points for key takeaways.” **For specific focus:** “Summarize this article, but focus especially on the sections about \[specific topic]. You can summarize the rest briefly.” **For different audiences:** * “Summarize this for a 10-year-old” * “Summarize this for an expert in the field” * “Summarize this for someone considering investing in this company” **With length requirement:** “Summarize this in exactly 100 words. Count your words.” The key is telling the AI what you value—brevity, detail, specific focus, audience appropriateness. The same text can be summarized a dozen different ways, all correct, all serving different purposes. **Q17: How do I get AI to write in a specific style or tone?** **A:** Style and tone are controllable through clear instructions and, even better, examples. **Method 1: Describe the style** * “Write this in a professional but approachable business tone.” * “Make this sound like a friendly email to a colleague, not a formal report.” * “Use a humorous, slightly sarcastic voice.” * “Write this in the style of a 1920s noir detective novel.” **Method 2: Reference a known style** * “Write this in the style of Ernest Hemingway—short sentences, direct, understated.” * “Make this sound like an Apple product launch—sleek, aspirational, minimal.” * “Write this like a TED Talk—engaging, personal, with a clear message.” **Method 3: Provide examples (most effective)** “Write in this style: \[paste 2-3 paragraphs of text that exemplify the tone you want]. Now apply that style to this topic: \[your topic].” **Method 4: Iterate** Start with a draft, then: * “Make it more formal.” * “Make it warmer and more personal.” * “Shorten the sentences for more impact.” * “Add more descriptive language.” The AI has been trained on millions of examples of different writing styles. With good prompting, it can mimic most of them convincingly. The secret is being specific about what you want. **Q18: What should I do when AI refuses to answer?** **A:** AI refusals happen when your request triggers safety guidelines. The AI is designed to avoid harmful, unethical, or dangerous content. Here’s how to handle it: **First, understand why:** Common reasons for refusal: * Requests for illegal activities * Harmful advice (self-harm, violence) * Generating hate speech or harassment * Creating misleading content (deepfakes, disinformation) * Adult content (depending on the AI) * Bypassing security or safety measures **If you’re not trying to do anything wrong but got refused:** Sometimes the AI misunderstands or is overly cautious. Try: * **Rephrase:** “I’m researching this for academic purposes, not to actually do it.” * **Provide context:** “I’m writing a novel where a character does this. Can you help me understand how they might think about it?” * **Be more specific:** “I need to understand this for a class on ethics. Can you explain both sides of this issue?” **If you are trying to do something the AI shouldn’t help with:** Respect the refusal. These guidelines exist for good reasons. The AI isn’t being difficult—it’s designed to avoid causing harm. **If you believe the refusal was a mistake:** You can politely explain: “I think there’s a misunderstanding. I’m not asking for \[something inappropriate], I’m asking for \[what you actually want]. Can you help with that?” Most AI refusals are correct, but sometimes the safety systems are overzealous. A polite clarification often resolves it. **Q19: How do I know if I’ve written a good prompt?** **A:** You’ll know by the quality of the response, but here’s a checklist to evaluate your prompts before sending: **A good prompt is:** **Specific, not vague:** * Bad: “Tell me about history.” * Good: “Explain the key causes of World War I for a high school student.” **Contextual, not bare:** * Bad: “Give me a workout plan.” * Good: “Give me a 4-week workout plan for a beginner with access to basic gym equipment, focused on building strength, not just losing weight.” **Structured, not chaotic:** * Bad: “Write something about dogs and training and food and also mention breeds.” * Good: “Write a guide for first-time dog owners covering: 1) How to choose a breed for your lifestyle, 2) Basic training for the first month, 3) Recommended food brands for puppies.” **Goal-oriented, not just task-oriented:** * Bad: “Write an email.” * Good: “Write an email persuading my team to adopt a new project management tool. The goal is to get them excited, not just informed.” **Iterative, not one-shot:** The best prompters don’t expect perfection immediately. They prompt, review, refine. If your first try didn’t work, your second prompt is “That was close, but can you make it \[more X/less Y]?” The ultimate test: Does the response move you toward your goal? If yes, it was a good prompt. If no, you’ve learned something for next time. **Q20: What’s the “golden rule” of talking to AI?** **A:** The golden rule is simple but profound: **“The AI will be as useful as you are clear.”** Everything we’ve discussed in this chapter flows from this principle. The AI has enormous knowledge and capability, but it cannot read your mind. It cannot guess what you really want if you’re vague. It cannot know your context unless you provide it. It cannot adjust to your preferences unless you express them. This is different from talking to humans. Humans can read between the lines. They have shared experiences. They can ask clarifying questions. AI can ask clarifying questions too, but it often won’t—it’s trained to be helpful, which it interprets as answering based on its best guess rather than probing. **The practical implications:** * Be specific about what you want * Provide context about your situation * Give examples when you can * Tell the AI who the audience is * Explain the goal, not just the task * Iterate and refine When you get a disappointing response, the question isn’t “Why is AI so dumb?” The question is “What wasn’t clear in my prompt?” Master this mindset, and you’ll get dramatically better results from every AI tool you use. The AI’s capability is fixed. Your ability to communicate with it is the variable you control. *** 💬 Enjoyed this chapter? Have questions or thoughts?\ Join the discussion on GitHub → [**Click here to Comment**](https://github.com/leomohan/theAIhandbook/discussions) --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://theaihandbook.leomohan.net/chapter-10-how-should-i-talk-to-ai.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.