Ch.3: Why ChatGPT Feels Useless on the First Try

0:00 Welcome to Learning Podcasts. ChatGPT for Everyone: Your 1st Real Conversation. Today we look at the gap between getting a useless answer out of ChatGPT and getting one you can actually send. The funny part is, most of the time the model is fine. The conversation around it is where it falls apart. Here is a scenario that plays out a thousand times a day. Two people open ChatGPT a minute apart. The 1st types, help me write an email. They get back a polite, generic, completely useless draft. They read the 1st line, sigh, and close the tab.

0:34 And that is the moment where most people give up on the tool entirely. They walk away thinking it is overrated. The 2nd person types something longer. I need to write an email to my landlord about a broken heater. The heater stopped working 3 days ago, I have already called the building line twice, and I want to be firm but polite because I plan to renew my lease. And that one comes back as something they could send without changing a word. Same model. Completely different room they walked into.

1:02 ChatGPT did not get smarter between those two attempts. The person just gave it something real to work with. Most people use ChatGPT the way they use Google. Type a question, read the answer, close the tab. That approach works. It also leaves about 80 percent of the value on the table. I will be honest, I did this for months. Type, read, close. It is the path of least resistance, because it is the muscle memory we already have. But hold on. For a quick factual lookup, the one-shot Google pattern is actually fine, right?

1:36 Like, what year did something happen, when does daylight saving time end, that kind of thing. For pure factual lookup, sure, one shot is fine. The trap is when people use the same one-shot pattern for tasks that are not factual lookup. Writing, planning, brainstorming, editing. Those are conversations. The single-question habit just carries over because it is what we already know how to do. Picture someone walking into a tailor's shop and saying, make me a suit. Then walking out before the tailor can ask a single question.

2:10 What color, what fabric, is it for a wedding or a job interview. The tailor is just standing there with the tape measure. They could produce something. It would be generic at best and wrong at worst. Yeah. And that is what a single message to ChatGPT looks like. You handed it nothing to work with, and then you graded the result. The model is not refusing to help. It is averaging across every possible request that fit your three-word prompt. And the average of every possible request is, by definition, generic.

2:46 Right. You did not ask for your suit. You asked for a suit. Right. So how much of this context are we actually supposed to hand over. Think of context the way you would think about ingredients in a recipe. A chef with salt, pepper, and a glass of water can make very little. Yeah, that is going to be a rough dinner. Hand that same chef vegetables, protein, spices, and a good stockpot, and the possibilities open up. Same chef. Different result. ChatGPT works the same way, then. It is not that the model magically gets smarter when you write more.

3:20 Exactly. You finally told it which version of the answer you actually wanted. This does not mean you have to write a novel every time you open a conversation. It just means spending, you know, 15 seconds thinking about what ChatGPT would need to know to give you something useful. So what does that look like in practice. Four questions. Who is the audience. What tone do you want. What have you already tried. Are there constraints, like a word limit or a deadline. That is it. 15 seconds, four questions, and most of the difference is already made.

3:53 Once ChatGPT gives you a response, resist two reactions. Do not accept it whole. Do not throw it out. Read it like a rough draft from a colleague. What did it get right, what missed. So the 1st message is not the answer. It is the starting point. Treating the 1st response as finished is one of the most common mistakes new users make. I mean, I have done that myself. You ask, you scan the answer, you go, eh, this tool is bad, and you close the tab. Yeah. Almost nothing ChatGPT writes for the 1st time is what you actually wanted to send.

4:28 It is close. It is rough. The work is in reading carefully, deciding what to keep, and asking for the next version. Take the landlord email. Maybe the structure is right but the tone is too formal. Maybe it included a detail you did not mention and do not want. Maybe the opening paragraph is perfect but the closing is weak. Right. All of that is useful information. Because it tells you what to ask for next. The gap between what you got and what you actually need is not a problem. It is the next prompt.

5:03 Which means the 2nd message you send is, kind of, more important than the 1st one ever was. Well, hold on. There is a whole school of thought, prompt engineering, that is the opposite of what you are describing. Spend serious time crafting one really good 1st prompt, with examples and constraints, and the model gets it right the 1st time. Why not just do that. For repeatable work, that is exactly the right move. If you are summarizing a hundred customer reviews, or extracting structured data from a stack of invoices, you build the prompt once, you tune it, you reuse it.

5:38 The investment pays off because you run it a hundred times. But the landlord email is a one-time thing. Right. For a one-time task, a 15-2nd context plus three short follow-ups is faster than spending 10 minutes on the perfect prompt. The deciding factor is, basically, are you going to use this prompt again, or just trying to get one specific thing done. So one-shot prompting and iteration are not really in competition. They solve different problems. Exactly. And the skill that pays off across both is the same one.

6:09 Read the response carefully, notice the gap, adjust. Whether you are tuning a saved template you will run a hundred more times, or sending a 3rd follow-up in the landlord chat, you are doing the same thing. The underlying habit is identical. So what does steering actually look like once you have a 1st draft on the screen. After reading the 1st draft, you can say something like, make the tone less formal, more like how I would actually talk to someone in the hallway. Or, remove the part about legal obligations, I do not want to threaten anything.

6:40 Or, keep the 1st two paragraphs but rewrite the last one to end with a clear request for a specific repair date. Each follow-up narrows the gap. The conversation itself is how you get there. It is the same shape as a haircut, actually. The 1st cut gets you in the right neighborhood. Then it is, a little shorter on the sides. Then, the bangs need to come up half an inch. Yeah, each correction is smaller and more specific than the last, and you converge on what you wanted in three or four passes. And you do not expect the 1st cut to be the final cut.

7:14 The conversation works the same way. And steering is not the only move. Sometimes you do not want a smaller change to the same draft. You want a completely different shape of the same idea. You can also ask for the same thing from a completely different angle. Give me a version that is much shorter, just three sentences. Or, now write one that is more emotional. Or, take the same content and make it sound like a quick voice note instead of a formal letter. Right. Seeing variations helps even when none of them is exactly right on its own.

7:47 You can usually point at one and say, the closing of this one, with the tone of that one. That single sentence gets you closer than another generic prompt would. Well, sometimes a conversation drifts. You started on the landlord email, then you asked about tenant rights, then something completely unrelated. And here is the part that is not obvious. ChatGPT keeps everything you have said in that conversation as context, so the model is still trying to account for the tangents when you come back to the email.

8:20 So it is, kind of, a whiteboard you keep writing on without erasing. Yeah, that is exactly it. By the time the board is covered, the new note you just added is competing with 20 old ones for the model's attention. The answers slip, not because the model got worse. The room got noisier. And starting a fresh chat is one click. And it is free. Right. A good rule of thumb is to start fresh whenever you switch to a genuinely different task. The accumulated context is an asset when it is relevant and a liability when it is not.

8:51 Well, if you want to feel the difference, the only way is to try it on something real. A tricky email you have been putting off. A cover letter you owe someone. A difficult conversation you have been rehearsing in your head. Yeah. Pick one. Do not invent a test prompt. Right. Spend a few seconds on the four questions before you type. Audience, tone, what you have already tried, what you need to avoid. Then read what comes back like a rough draft, and ask for one specific change. Do it again. By the 3rd or 4th round, you will likely have something you can actually use.

9:25 The point is not that every task needs four rounds. The point is that four rounds is often available to you, for free, in under 2 minutes. Most people close the tab after the 1st response and never find out. So this loop is not just a thing for landlord emails. The landlord email is the example, but the loop is the lesson. Imagine you are a parent with a parent-teacher conference tomorrow morning. You want a short list of talking points, because the meeting is 15 minutes long. And you do not want to forget the important questions the moment you sit down in that tiny chair.

10:01 You give ChatGPT the situation the same way you gave it the broken heater. Your child is in 2nd grade, reading well above grade level, quiet during class discussions, going through a friendship shift. You want to be efficient. Then you read the response, notice that one of the suggested questions feels too confrontational, ask for a softer version. Yeah. Add a new detail mid-conversation about a new sibling at home, ask for a printable list. By the 4th message, you have something you can take into the meeting.

10:31 So the 1st message is not the answer. It is the start of a conversation, and the people who get the most out of ChatGPT are the ones who keep going. It is not about asking the perfect question. It is about starting somewhere reasonable and refining from there. Next chapter, we look at how to phrase the request itself so the very 1st response comes back closer to what you actually wanted. Thanks for listening to Learning Podcasts.