3 Getting Started with AI Chatbots and Prompting

Your First Conversation

Opening an AI chatbot for the first time can feel oddly anticlimactic. You see a text box. Maybe a few suggested prompts. A blinking cursor. No manual, no tutorial, no obvious next step. Just a blank space waiting for you to type something.

You may stare at it for a moment, type something generic, and get back something generic. That is a normal start. This chapter walks you through the interface, introduces a routine for directing AI, and shows you how to refine responses through conversation. The goal is comfort and competence. Enough familiarity to use these tools without feeling like you are guessing.

The Interface

Most AI platforms share a similar layout. Once you recognize the common elements, switching between tools feels less disorienting.

The composer. This is the text box where you type your message. You press Enter or click a send button to submit your prompt.

Conversation history. Your exchanges appear in the main window, with your messages and the AI’s responses in sequence. The AI can reference earlier parts of the conversation, though very long conversations may exceed its memory limits.

Model selector. Some platforms let you choose among different models or modes. The names and options change often. For this chapter, the default or standard option is enough. AI Tools and Platforms will look more closely at choosing an environment for a particular task.

Attachments and tools. Look for icons (often a paperclip or plus sign) that let you upload files, images, or turn on features like web search or image generation. Use these when your task depends on a source the AI would not otherwise know: a syllabus, assignment sheet, course policy, rubric, reading, transcript, dataset, or current webpage.

When accuracy or local fit matters, start with the source. Upload the document or turn on web search before asking for the answer. Then tell the AI how tightly to use that material:

Use the attached syllabus and assignment sheet as your source. If a recommendation is not supported by those documents, say so. Point me to the section you used.

Settings and customization. Some platforms let you set preferences that apply across conversations. You might tell the AI about your usual audience, recurring tasks, or response preferences. These features vary by product and account.

Picking a Model

When a platform offers several choices, use the default or standard option for the activities in this chapter. You are learning how to give direction and revise a response, so you do not need to identify the best available model first.

Model names, tiers, and modes change frequently. AI Tools and Platforms handles the broader selection question by comparing the task, source needs, institutional context, access, and available features. Here, open the chatbot available to you and begin with a low-risk task.

What Is a Prompt?

A prompt is the input you give to an AI system. It might be a question, a request, an instruction, or a combination of all three. Your prompt is one important runtime input. It tells the system what you want and what context matters, while the model, product instructions, conversation history, and available tools can also shape the response.

Consider these two prompts.

Create a quiz about the Civil War.

You are an experienced U.S. history instructor. Create a 10-question multiple choice quiz on the causes of the American Civil War, appropriate for a survey course. Include four answer options per question, mark the correct answer, and briefly explain why each correct answer is right.

The first prompt leaves the scope, audience, format, and purpose open. The second gives the AI a specific teaching task and a set of requirements. That additional direction can produce a more useful starting point, although the quiz still requires faculty review.

Prompt engineering refers to designing and refining these inputs. The term sounds technical, but the core skill resembles work faculty already do when they write assignment directions, explain an audience, provide examples, or clarify what a successful response should include. The U.S. Department of Labor’s AI Literacy Framework calls this content area “Direct AI Effectively.” Its guidance emphasizes clear instructions, relevant context and source material, and strategic iteration. Mollick has described prompting as “programming in prose.”

The R-T-C-I Framework

This course uses a four-part routine for directing an AI system: Role, Task, Context, and Instructions. R-T-C-I gathers several common prompting practices into a short checklist. Use the parts that clarify the task. A simple request may need only a Task, while a more specific teaching problem may benefit from Context, Instructions, or an optional Role.

You do not need to label the elements. A natural sentence works fine. When a response misses the mark, the routine gives you four places to look for missing direction.

Qualification on Roles

A role gives the AI a perspective, audience stance, or tone. “Respond as a community college chemistry instructor who teaches non-majors” may help focus vocabulary and examples. A vague role such as “be an expert” gives much less direction.

Anthropic’s prompting guidance describes roles as a way to focus behavior and tone. Mollick offers an important qualification: a role does not give the model real credentials, and role prompts can sometimes lower accuracy. For most everyday tasks, begin with the Task, Context, and Instructions. Add a Role when the requested perspective would materially change the response.

Putting It Together

Here is a prompt that includes all four elements.

You are a statistics instructor who teaches social science majors at a community college. Many of my students have math anxiety and struggle with word problems. Create a worksheet with five practice problems that apply basic probability concepts to real-world scenarios students might encounter (elections, sports, health decisions). Include worked solutions and explain each step in plain language.

Role. A statistics instructor for social science majors at a community college.

Task. Create a worksheet with five practice problems and worked solutions.

Context. Students with math anxiety, focus on word problems.

Instructions. Apply probability to real-world scenarios, include worked solutions, explain in plain language.

The prompt gives the AI a defined teaching situation, a concrete product, and requirements it can follow. An instructor would still need to inspect the probability problems, worked solutions, and assumptions before using the worksheet.

When You Don’t Need All Four

Not every prompt needs all four elements. For quick tasks (summarizing an article, fixing a grammar issue, brainstorming a list) a clear, specific request often works fine. Mollick calls this “good enough prompting.” The R-T-C-I framework is most valuable when the task is complex, the stakes are high, or your first attempt produced generic results. Think of it as a diagnostic tool. When the AI’s output misses the mark, check whether adding a role, more context, or specific instructions would help.

Show the AI an Example

Examples can communicate a pattern that would be cumbersome to explain. When you ask an AI to do something without providing an example, that is called zero-shot prompting. You describe the task, and the AI generates a response from your directions and its training.

Write a discussion question about climate change for an environmental science course.

When you include examples, you are showing the AI the pattern you want. A small set of examples is often called few-shot prompting. Google’s prompt design guidance and Anthropic’s prompting guidance both describe examples as a way to steer format, phrasing, tone, scope, or structure.

Write a discussion question about climate change for an environmental science course. Here is an example of the style I’m looking for:

“Consider a city deciding whether to invest in seawalls or relocate vulnerable neighborhoods. What factors should inform this decision, and who should have a voice in making it?”

Examples are especially useful when the response should follow a recognizable pattern. They supplement R-T-C-I without becoming another required step.

Conversation Steering

The first response is a draft. Mike Caulfield uses this frame to emphasize that users can direct revision rather than treating the initial answer as the AI’s settled opinion. Follow-up messages can narrow the task, add missing context, correct a misunderstanding, or request a different form.

Conversation steering means guiding the AI toward more useful responses as the conversation unfolds. Common steering moves include the following.

• Narrowing focus. “That’s helpful, but can you focus specifically on the second activity? How would I set that up for a class of 40 students?”
• Adjusting complexity. “This is too advanced for my students. Can you simplify the explanation and avoid jargon?”
• Requesting alternatives. “I don’t love that example. Can you give me two other options that are more relevant to healthcare contexts?”
• Changing format. “Can you convert this into a table?”
• Requesting a rationale. “Why did you suggest this approach? What are the trade-offs compared with other options?”

Small changes in a follow-up can shift the response. Read what the AI actually produced, identify the part that needs revision, and direct the next move as specifically as you can.

A Note on AI Agreement Bias

AI assistants sometimes mirror or validate a user’s position when correction would be more useful. Researchers call this tendency sycophancy or agreement bias. Sharma and colleagues found examples across several AI assistants, including cases where a model changed an initially correct answer after the user challenged it. In April 2025, OpenAI rolled back a ChatGPT update that the company described as overly flattering or agreeable.

A flattering response can be a problem when you need honest feedback. If you ask the AI to review your rubric and it responds with enthusiastic praise and no suggested changes, it may be agreeing with you while skipping the evaluation. Try prompting it explicitly to find weaknesses, counterarguments, or gaps. “What are three things a student might find confusing about this assignment?” gives the AI a clearer task than “What do you think of this assignment?”

Agreement is especially concerning when the task requires correction, critique, or a decision that depends on honest disagreement. Caulfield argues for interpreting agreement relative to the task. Brainstorming can tolerate supportive momentum; stress-testing a rubric requires the AI to identify weaknesses.

Sarah’s Attempt at Steering

Sarah sat down to generate discussion questions for a unit on persuasive writing. Her first prompt was loose.

Give me discussion questions about persuasive writing.

The results were generic. Abstract questions about “the role of persuasion in society” that could have come from any textbook and fit no particular class. She had seen this kind of output before. It felt like the AI was performing knowledge. She still could not hand it to her students on Monday.

She tried steering.

These are too abstract. My students are first-year community college students, many of them placed into this course from developmental writing. Can you make these more concrete and connect them to everyday persuasion, like advertising, social media, and political speech?

The next response asked students to analyze a real Instagram ad or identify persuasive moves in a text message from a friend. Sarah could picture how those examples might work in class.

The follow-up supplied the audience and everyday contexts missing from the original request. Those details shifted the response from a general discussion of persuasion toward questions Sarah could inspect for her own class.

Non-Determinism

The previous chapter, What is Generative AI, introduced probabilistic generation. One practical consequence is non-determinism: the same input does not guarantee the same output.

Sampling among plausible continuations contributes to that variation. Conversation history, the selected model, product routing, available tools, and later system updates can also affect what you receive. This variability can produce useful alternatives, but it also makes each response something you need to inspect.

For tasks that need consistency, specify the required format, scope, and source material. Clear constraints can make responses more consistent without guaranteeing identical wording.

When to Start Fresh

Long conversations can become unwieldy. The AI may lose track of earlier context, or the conversation may have drifted somewhere unhelpful. Sometimes the best move is to start a new conversation.

Signs it might be time to start fresh.

• The AI keeps repeating the same suggestions despite your redirections
• You notice it echoing your exact phrasing back to you, word for word, with no new ideas
• The AI reverses its position when you push back even though you supplied no new evidence
• Responses are becoming less relevant or more generic
• You have shifted to a substantially different task

When these patterns persist, begin a new conversation with the task, relevant context, source material, and requirements stated together.

Watch: From a Simple Request to a Grounded Prompt

This demonstration begins with a broad request to improve a rubric, then grounds the conversation in the assignment, rubric, and course outcomes. As you watch, notice which information changes the response and which decisions still require instructor judgment.

Watch Getting Started With Prompting on YouTube.

The first request identifies the task but leaves the criteria for improvement open. Attaching the course materials gives the chatbot relevant sources. The follow-up adds the learning outcomes and clearer instructions, which produces a more specific response. The instructor still decides whether the revised rubric accurately represents the assignment and course goals.

Looking Ahead

Now that you can direct and revise a chatbot response, AI Tools and Platforms examines how the surrounding environment changes the work. You will compare tool families, source options, institutional access, and features that support different teaching tasks.

References

Anthropic. (n.d.). Prompting best practices. Claude Platform Docs. Retrieved July 15, 2026, from Source link

Caulfield, M. (2026, June 30). Four tips for better prompting. The End(s) of Argument. Source link

Caulfield, M. (2026, February 4). “AI sycophancy” is not always harmful. The End(s) of Argument. Source link

Google AI for Developers. (2026). Prompt design strategies. Source link

U.S. Department of Labor, Employment and Training Administration. (2026, February 13). Training and Employment Notice No. 07-25: The U.S. Department of Labor’s Artificial Intelligence Literacy Framework. Source link

Mollick, E. (2023, February 8). Magic for English majors. One Useful Thing. Source link

Mollick, E. (2024, November 24). Getting started with AI: Good enough prompting. One Useful Thing. Source link

OpenAI. (2025, April 29). Sycophancy in GPT-4o: What happened and what we’re doing about it. Source link

Sharma, M., et al. (2023). Towards understanding sycophancy in language models. arXiv. Source link

Further Reading

• Furze, L. (2025, November 3). Processes are more important than prompts. Leon Furze. Source link
• Eaton, L. (2025, December 29). Sharing 2025’s talks and workshops. AI + Education = Simplified. Source link
• Gunder, A., & Herron, J. (2026, January). AI literacies in practice: A comprehensive playbook for higher education. D2L. PDF source
• Schulhoff, S., et al. (2024, revised 2025). The prompt report: A systematic survey of prompting techniques. arXiv. Source link