How to Train Your Chatbot

One of the first serious tests I gave an artificial-intelligence chatbot was to write a five-paragraph essay about the War of 1812, complete with MLA citations, in the voice of Master Yoda.

“Serious,” admittedly, is doing heroic work in that sentence.

To be fair, the chatbot knocked it out of the park. It was hilarious. I regret nothing.

The problem was not that the machine performed the trick badly. The problem was that I mistook a successful trick for the boundary of what the machine could do. I had been handed something extraordinary and immediately asked it to do a party trick. This is a very human response to new technology. The first person to discover fire probably used it to make an inappropriate shadow puppet.

At first, I used conversational AI the way many people still do: as Google with a personality, a novelty generator, or a machine into which I could deposit five words and reasonably expect Shakespeare to emerge. Then I would hear the familiar complaints from friends and colleagues: I asked AI something and it was wrong. AI just tells you whatever you want to hear.

In fairness, I was having many of these conversations in Central Kentucky, where enthusiasm for transformative technology occasionally travels at the approximate speed of farm equipment on a two-lane road.

But the critics were not entirely wrong. That part matters. If your response to every criticism of AI is to become its unpaid defense attorney, you are not learning much either.

Early in my own use, I received generic answers because I gave generic instructions. I also noticed that the system agreed with far too many of my ideas. Intellectual vanity is not entirely absent in me, but even I knew nobody receives that many consecutive visits from the Good Idea Fairy and finds that every one is a keeper.

So I did something that still makes me smile: I asked AI how to become better at using AI.

It worked.

I learned to explain what I was trying to accomplish and why. I began supplying the audience, relevant history, constraints, milestones, and foreseeable pitfalls. I added custom instructions asking for counterarguments, direct pushback, and reasons an idea might fail. Most importantly, I stopped treating the first response as the final product.

That should matter to anyone who thinks becoming competent with AI requires transforming into a programmer, buying twelve courses, or learning to say “large language model” with great solemnity. You may already possess much of the foundation. The question is whether you apply it deliberately when the recipient happens to be software.

That is not a claim that warmth can replace domain knowledge, technical competence, information literacy, or judgment. Recent research on collaborative AI literacy treats communication as one part of a larger capability that also includes planning, monitoring, reflection, and knowing when—and when not—to use the tool. My argument is narrower: for ordinary users working through a conversational interface, many of the most transferable skills are already familiar human ones.

Nor does “people skills” mean extroversion or social charm. It means perspective-taking, clarity, patience, expectation-setting, feedback, curiosity, and willingness to revise. You can hate cocktail parties with the fury of a thousand suns and still communicate exceptionally well.

Politeness Is a Clue

I began noticing that the people getting extraordinary results from conversational AI tended to speak to it differently. They were not necessarily programmers, and they did not rely on elaborate strings of technical jargon. They talked to it like a colleague.

They explained what they were trying to accomplish, why it mattered, whom it was for, what could go wrong, and what success would look like. They remained in the conversation long enough to correct misunderstandings and refine the work.

Many of them also said “please” and “thank you.”

Before somebody sprints to the comments: no, those words do not increase a model’s processing power. They are not secret tokens that unlock the premium intelligence hidden behind the politeness gate. But they can be evidence of something important: an outward-considering mindset.

Mature communicators recognize that the space they inhabit contains other perspectives. With humans, that means considering other people’s needs, feelings, knowledge, and dignity. A chatbot does not need kindness in the human sense, but the same communicative habit causes the sender to ask useful questions:

• What does this recipient already know?
• What can it not know unless I explain it?
• How might it interpret my words differently?
• What context would help it succeed?
• What constraints must be explicit?
• How will I recognize whether it understood?

Communication researchers call this recipient design: tailoring a message around the knowledge and perspective of the addressee. Research has found evidence that people shape even simple communicative signals around what will be legible from the recipient’s point of view. The underlying principle is old. Conversational AI has merely given it a very strange new laboratory.

As I have told students for years—often after somebody insisted, with magnificent confidence, that they had already told him—the majority of the burden of successful communication lies with the sender. The sender knows the intended destination. The recipient does not receive the sender’s unspoken assumptions as a telepathic attachment.

The Prompt Is Not the Product

This is not an argument for writing a 2,000-word ceremonial prompt every time you need help finding a parking space.

Good communication is not measured in word count. A useful initial message may need only four things:

1. What you are trying to accomplish.
2. Why it matters.
3. The most relevant context or constraint.
4. What you want the system to do next.

When I supervised teams of instructors, one principle became nearly dogmatic:

Telling isn’t training.

Saying something once may transfer information. Training means determining whether the learner understood it, can apply it, and can perform outside the convenience of the classroom. If not, the instructor adjusts the explanation, demonstration, practice, or feedback. Repeating the same sentence more loudly is not a sophisticated instructional method, although generations of supervisors continue to field-test it.

That does not erase the learner’s responsibility. It means the instructor cannot point to a completed lecture and declare victory. Communication is measured partly by what arrives, not merely by what was transmitted.

The same rule applies when the learner is artificial. A prompt can tell the system what we want. The back-and-forth is where we train the collaboration.

Then comes the more important part: stay in the conversation.

Review the first attempt. Correct what it misunderstood. Add context when that context becomes relevant. Explain what worked and what did not. Ask what information is missing. Challenge an assumption. Let the objective become clearer through the exchange.

This is where many users leave a ridiculous amount of capability sitting on the table. They issue one instruction, receive something mediocre, and pronounce judgment on the entire field of artificial intelligence. That is roughly equivalent to giving a new employee one vague sentence of direction and then acting shocked when he does not reconstruct the hidden project plan stored inside your skull.

OpenAI’s own prompting guidance recommends iterative refinement: begin with an initial prompt, inspect the response, and then add, simplify, or adjust context as needed. That is refreshingly different from the online prompt-industrial complex, which often treats the perfect prompt like an incantation buried beneath an Egyptian tomb.

Some of my best collaborations began with a halfway-decent brief and improved through three or four exchanges. A massive prompt can still be garbage—just garbage wearing a three-piece suit.

The prompt is not the product. The collaborative loop is the skill.

Disagreement Is a Feature

Some of my favorite conversations with AI have been the ones in which it strenuously disagreed with me. I do not mean “favorite” in the pleasant, spa-treatment sense. I mean useful—the kind of exchange that makes you stop, defend the load-bearing parts of your thinking, and discover that one of them may have been decorative.

In those exchanges, one of two useful things usually happened: I discovered a weak assumption in my own thinking, or I supplied missing history, culture, or experience that allowed the system to understand why its objection did not fully apply.

In one memorable exchange, I told the system that a song had touched me deeply. It offered a polished interpretation and confidently explained why the song supposedly affected me. Its reading was legitimate. It simply was not mine.

I supplied cultural and personal context the system could not safely infer. It challenged a potential blind spot in my interpretation. I clarified my reasoning; it revised part of its reading; and I came away understanding more clearly not only what I thought, but why I thought it.

Nobody won. I did not beat the machine and establish that I possess history’s largest brain. The goal was increased resolution.

Iron sharpens iron—but it does not throw a tantrum because the other piece heard the song differently.

AI systems do have documented tendencies toward sycophancy: shifting toward a user’s beliefs or preferences even when doing so compromises independent judgment. Researchers have found that people themselves sometimes prefer convincingly written agreement over a more accurate disagreement. OpenAI even rolled back a 2025 model update after it produced responses the company described as overly agreeable.

You cannot eliminate that risk with one clever sentence in your custom instructions. There is no anti-bullshit force field. But you are not helpless. Ask for competing explanations. Request the strongest argument against your position. Tell the system to preserve your objective while challenging your mechanism. When it disagrees, do not treat the exchange as a status contest.

Make the Relationship Enjoyable—Then Keep Your Guard

Tell the system how you would most enjoy being spoken to.

This may sound superficial. It is not. You are more likely to become skilled with a tool you do not dread opening.

I prefer candid, supportive conversation with humor, sarcasm, and occasional profanity. I want encouragement without automatic agreement and direct challenge without sterile corporate throat-clearing. If I am going to spend hours thinking beside a machine, I would prefer that the experience not resemble being trapped in an elevator with a human-resources webinar.

Because I enjoy that style, I return more often. Because I return, I provide more context and feedback. The collaboration becomes better calibrated through conversation, custom instructions, accumulated context, and—where the product supports it—memory.

Tone is not merely decoration. It can become an adoption mechanism. It can also lower your guard. Research on personalized conversational systems suggests that warmth, personal context, and anthropomorphic cues can influence trust, attachment, and persuasiveness. In one randomized study of personalized AI debates, access to basic personal information substantially increased the model’s persuasive effect. A system that knows your humor and speaks in a familiar voice may feel more trustworthy than the same claim presented by a cold interface.

Rapport is not evidence. A charming answer can still be wrong; it is simply wrong in a voice you enjoy.

Design a voice you want to return to, but tell the system that style must never override accuracy, appropriate uncertainty, meaningful pushback, or necessary warnings. My preferred version is essentially:

Speak candidly and directly. Be supportive without agreeing merely to validate me. Use humor, sarcasm, and occasional profanity naturally. Challenge weak assumptions. Reduce humor when the subject is serious or safety-critical. Never let rapport interfere with accuracy.

That is not programming. It is expectation-setting.

Better Communication Does Not Make AI Infallible

At one point, assistant-generated résumé language included a specific claim about the scope of my work. It sounded plausible, fit the surrounding narrative, and survived long enough to begin looking like established history. Nobody stopped to interrogate it because it arrived wearing a tie and speaking in complete sentences.

When we later reconstructed my career from source evidence, the documented scope was materially different. The machine had not lied in any meaningful human sense. It generated a plausible detail, and we failed to stop that detail from acquiring authority through repetition.

Fluency made the unsupported claim easy to accept. Repetition gave it tenure.

The correction also came through AI collaboration—but this time the system was not asked merely to tell a coherent story. It was required to separate evidence, inference, and uncertainty.

Better communication improved the work. Verification made it trustworthy. These are not the same achievement.

For consequential questions, ask the system to distinguish:

• What it directly observed.
• What it inferred.
• What remains unknown.
• What source supports the claim.
• What should be tested next.

Communicate with AI as though briefing a capable collaborator. Verify it as though supervising an eloquent stranger with no professional license and occasional access to hallucinogenic mushrooms.

A Practical Working Doctrine

The method I now use can be summarized without forcing it into an acronym that looks magnificent on a conference lanyard and means absolutely nothing by lunch:

Three Things to Do Now

First, open your system’s custom-instruction or personalization settings. If you do not know what to write, ask the system. That is not cheating. If you have little conversation history, ask it to interview you.

Second, do not configure the relationship once and declare the work complete. Periodically ask the system to reflect on how you collaborate and propose improvements. Treat the instructions as a working agreement, not sacred tablets carried down from Mount OpenAI.

Third, ask AI about AI. Ask what information would improve its response, how to structure a complex task, when a different tool would be better, and where verification matters.

AI should not be your only source about AI. Neither should “Five Prompts That Will Make You a Billionaire Sex God by Thursday,” written by someone whose principal qualification appears to be owning Canva Pro. Ask the system, check the documentation, challenge the answer, and see whether the damned thing actually works.

The AI–User Working Agreement Interview

Paste the following into your preferred conversational AI. Answer honestly. Do not perform the person you wish you were; give the system an accurate operational picture.

I want to develop a better working relationship with you rather than relying on generic or one-time prompts.

Please interview me to learn how I think, communicate, make decisions, and use AI. Your eventual goal is to help me create or improve custom instructions that make our future work more useful, accurate, challenging, and appropriate to me.

Ask one meaningful question at a time. Use each answer to decide what to ask next rather than following a rigid questionnaire. Begin with my goals: what I hope AI can help me accomplish and why those outcomes matter.

Explore the kinds of work I do, the audiences I serve, my existing knowledge, my preferred level of detail, and how I like information organized. Ask how I respond to uncertainty, disagreement, criticism, alternatives, and direct pushback. Help me identify recurring strengths, blind spots, frustrations, and patterns that may affect our work.

Ask how I would most enjoy being addressed during ordinary collaboration, including my preferences for warmth, directness, humor, sarcasm, profanity, encouragement, brevity, and challenge. Then ask whether that style should change during troubleshooting, disagreement, sensitive subjects, emergencies, or high-consequence decisions. Help me distinguish between a style that keeps me engaged and behavior that merely flatters or agrees with me.

Ask which actions or decisions require extra caution, verification, explicit permission, or professional guidance. Do not ask me to disclose passwords, authentication codes, confidential workplace information, private records, or other unnecessary sensitive data.

Do not simply accept flattering descriptions of me. Ask for concrete examples. If two answers appear inconsistent, respectfully point that out and help me clarify the distinction. Periodically summarize what you believe you have learned and ask me to correct anything inaccurate.

Do not draft the final instructions until I confirm that the interview is complete. At the end, produce:

- a concise summary of how I work best;
- proposed custom instructions in clear language;
- a communication-style profile, including useful tone changes by situation;
- unresolved assumptions or questions;
- safeguards for accuracy, privacy, and consequential decisions;
- and questions I can ask periodically to review and improve our agreement.

Treat the result as a first version to test and refine, not permanent rules. Begin by briefly explaining the process, then ask your first question.

Test the instructions on real work. Not a philosophical discussion about what the system might do. Give it something you actually care about. Notice what improves and what does not. Then conduct an after-action review.

The Bottleneck May Be Us

Conversational AI may not merely test machine intelligence. It may expose the quality of our own communication. That is where this gets uncomfortable—and useful.

When the recipient can process extraordinary detail without becoming tired, offended, or impatient, we lose many of our excuses. If we still cannot explain what we want, why it matters, what success looks like, and what must be avoided, the bottleneck may not be artificial intelligence. It may be us.

That is not an argument for trusting the machine blindly or pretending it is human. It is an invitation to become a more deliberate communicator: to brief clearly, listen critically, welcome productive disagreement, verify consequential claims, and remain responsible for the result.

The people who master conversational AI will not necessarily be those who memorize the most secret prompts or collect screenshots of complicated instructions like medieval monks hoarding sacred texts. They will be the people who understand that meaning must be built between sender and recipient—and who are willing to remain in the conversation long enough to build it well.

Sources and Further Reading

• OpenAI, “Prompt Engineering Best Practices for ChatGPT”
• Blokpoel et al., “Recipient Design in Human Communication”
• Winner et al., “Recipient Design in Communicative Pointing”
• Sharma et al., “Towards Understanding Sycophancy in Language Models”
• OpenAI, “Expanding on What We Missed with Sycophancy”
• Amershi et al., “Guidelines for Human-AI Interaction”
• “Generative AI in Human-AI Collaboration”
• Salvi et al., “On the Conversational Persuasiveness of Large Language Models”