How to Use ChatGPT: 7 Pro Tips for Prompt Engineering & Custom GPTs
I've read probably fifty guides on prompt engineering. Most of them are useless. They give you abstract frameworks. Acronyms. Nothing you can actually use Monday morning when you're staring at a blank chat box.
So I'm going to skip the theory and tell you what actually improved my results. These are the specific techniques I reach for when ChatGPT gives me garbage output. Which still happens all the time by the way. Even after two years of daily use. You'd think I'd have it figured out by now. Honestly, I haven't.
The Role Trick Is Real
Here's the technique I use more than any other. Start every prompt by telling the model who it is. Simple. Obvious once you hear it. But most people never do it.
"You are an experienced patent attorney." Then ask your legal question. "You are a Pulitzer-winning journalist." Then ask for article feedback. "You are a senior Python developer who reviews code for a living." Then paste in your buggy script. The pattern is always the same. Identity first. Then task.
The role assignment does something interesting. It activates a different distribution of responses in the model's training data. A prompt about investment strategy gets wildly different answers when it comes from "a financial advisor" versus "a day trader" versus "a personal finance blogger." Same underlying topic. Different perspective. Different quality. Kind of remarkable actually.
I tested this systematically once. Same question, ten different roles. The answers varied in depth, tone, usefulness, everything. The best roles were specific senior professionals with clear expertise boundaries. The worst? No role at all. The difference was honestly bigger than I expected.
Chain of Thought Actually Works
You might have heard about chain of thought prompting. Sounds fancy. All it means is asking the model to show its work. That's it.
Instead of "what's 24 times 37" you say "what's 24 times 37, walk me through each step of the calculation." The accuracy improvement on math and logic tasks is real. Not a small improvement. We're talking about going from wrong half the time to right most of the time. That's meaningful.
I use this for anything that requires reasoning. Debugging code, analyzing arguments, comparing options, solving word problems, and so on. The step by step format forces the model to be more systematic and less likely to jump to wrong conclusions. Something about making it articulate the intermediate steps catches errors that would otherwise slip through.
But there's a tradeoff. Chain of thought makes responses longer. Sometimes much longer. If you just need a quick answer, skip it. If accuracy matters, use it. You'll develop a sense for when each approach makes sense.
Give It Examples
This is called few shot prompting in the literature. I just call it showing the model what you want. Before you ask your actual question, give it two or three examples of the format or style you're looking for. It works better than you'd think.
Say you want product descriptions for an online store. Instead of describing the format in words, show it: "Here are three product descriptions in the style I want. Scented Candle: Hand poured soy wax with lavender essential oil. Burns for 40 hours. Ceramic Mug: Stoneware with reactive glaze. Dishwasher safe, 12oz capacity. Leather Notebook: Full grain leather cover, 192 lined pages, refillable. Now write ten more descriptions for these products: wool blanket, desk lamp, wooden cutting board..."
The model picks up the pattern instantly. Sentence structure, length, the types of details included. Your examples teach it faster than any written instructions could. It's like showing a new hire three samples of what you want versus handing them a twenty page style guide.
I keep a folder of good examples for different task types. Email templates, code documentation formats, blog post outlines, you get the idea. When I need that kind of output, I paste in the examples first. Works way better than describing what I want in abstract terms.
Negative Instructions Matter More Than You Think
Most prompting advice focuses on what to do. Tell the model its role. Give it context. Be specific. All good advice. But telling it what NOT to do is equally important. And almost nobody mentions it.
"Don't use buzzwords like synergy or disruptive." "Don't include any emojis." "Don't write more than three sentences per paragraph." "Don't start any sentence with However or Moreover." These negative constraints work surprisingly well.
The model respects these constraints better than positive instructions sometimes. If you say "use simple language" it might still sneak in some jargon. But if you say "do not use any technical terms" it usually complies completely. There's something about a clear prohibition that the model handles more reliably than a vague stylistic preference.
I add at least one negative instruction to almost every prompt now. It's become a habit I don't even think about. And it consistently improves the output in ways that positive instructions alone don't achieve.
Custom GPTs Are Basically Saved Prompts
People overcomplicate custom GPTs. At their simplest, they're just saved prompts with some extra features bolted on. You write instructions once. Reuse them forever. That's the core value proposition.
The real value isn't in the technology. It's in the consistency. Every time you use the custom GPT, it starts from the same instructions. Same tone. Same format. Same rules. No variation unless you want it. For repetitive tasks this is genuinely game changing.
I have custom GPTs for the tasks I do weekly. Newsletter drafts, social media captions, code reviews, meeting agendas, stuff like that. Each one took maybe ten minutes to set up. Each one has saved me many hours over the months since I built them.
Building one is straightforward. Go to the Explore section in ChatGPT. Click Create. Fill in the instructions field with exactly how you want it to behave. Be specific about format, tone, what to avoid. Upload example files if you have them. Test it with a few prompts. Refine based on what comes back wrong. You don't need to publish them or share them with anyone. Most of mine are private tools I built for myself.
Learn to Iterate Fast
The biggest skill in using ChatGPT isn't writing good prompts. It's iterating quickly. First response off target? Type "shorter" or "more casual" or "add statistics" and get a revision in three seconds. Still not right? Try again with different language. The feedback loop is almost instant.
People who treat each prompt like a one shot deal get frustrated and quit. People who treat it like a conversation get results. The difference is nothing more than willingness to ask for changes. Iterate. Always iterate.
I probably average three iterations per task. For important work, maybe eight or ten. Each iteration takes seconds. The cumulative improvement from round one to round eight is dramatic. You'd be surprised how much better the output gets when you just keep asking for tweaks.
FAQ
Q: What's the single most important thing to include in a prompt?
Context about who the output is for. Tell the model the audience and everything else flows from that. An explanation of quantum computing for a ten year old looks nothing like the same explanation for a physics graduate student. The audience determines vocabulary, depth, examples, structure, everything. If you only add one thing to your prompts, add who it's for.
Q: How long should my prompts be?
Longer than you think. A good prompt for a nontrivial task is usually three to five sentences. You need role, task, audience, format, and constraints. That takes space. If your prompt fits in a single sentence, you're probably missing at least two of those elements.
Q: Can I use the same custom GPT for different tasks?
Technically yes but I wouldn't recommend it. Each custom GPT should do one thing well. If you try to make it handle everything, it becomes no better than the default ChatGPT and you've defeated the whole purpose. Specialization is the point. Build one per task.