How to Use ChatGPT: Expert Tips for Prompts, Custom GPTs & API

I nearly quit using ChatGPT after the first week. Everything I asked it came back sounding like a robot wrote a textbook. Which I guess is exactly what happened.

What changed things for me was stumbling onto the idea of giving the model a specific persona to play. Not just "be helpful." But "you are a specific person with specific expertise and a specific way of talking." That one shift took my results from useless to useful in about ten minutes.

Here's everything I've figured out since then about getting ChatGPT to actually do what you want.

The Prompt Formula I Use Every Day

I've boiled it down to something I can remember without looking at notes. Four parts. Every prompt I write hits at least three of them.

Role. Tell it who it is. "You are an experienced product manager at a SaaS company." Not optional. This is the part that most dramatically changes output quality.

Task. What exactly do you need. Be painfully specific. "Write product requirement doc" is too vague. "Write a two page PRD for a mobile feature that lets users export their data as CSV" is specific enough to work with.

Audience. Who is this for. The same task written for executives looks different from the same task written for engineers. Tell the model who will read the output.

Constraints. Format, length, tone, things to avoid. "Under 300 words. Use plain English. Do not use the word leverage as a verb."

That's it. Most people skip at least two of these. Usually audience and constraints. And then they wonder why the output misses the mark.

Advanced Stuff That's Actually Worth Learning

After the basics, there's a handful of techniques that genuinely improve results for specific situations. Not everything labeled "advanced" is useful. Some of it is academic nonsense that sounds impressive but doesn't help in practice.

Chain of thought prompting is worth learning. It's the one where you ask the model to reason step by step instead of jumping to an answer. For math, logic, debugging, or any kind of analysis it makes a real difference. The cost is longer responses. Worth it when accuracy matters.

Few shot prompting is also worth learning. Give the model two or three examples of the output format you want before asking your actual question. The examples teach the pattern more efficiently than any description could.

Temperature control matters if you're using the API. Lower for facts and code. Higher for creative writing and brainstorming. The default is usually fine but knowing how to adjust it gives you more control.

System messages in the API let you set rules that persist for the entire conversation. Think of it as the constitution for that chat session. Everything the model does should comply with the system message. Use it for tone, behavior rules, and output format standards.

Stuff you can safely ignore for now includes fine tuning, function calling, embeddings, and most of the advanced API features. They're useful for specific use cases but not something a general user needs to learn.

The Desktop App Changed How I Use It

I mostly use the desktop application now instead of the browser. It has a global keyboard shortcut that brings up a chat window from anywhere. Option plus Space on Mac. I don't know the Windows equivalent.

This sounds like a minor convenience but it changes how often you reach for the tool. Instead of opening a browser tab and navigating to the site, you hit a key combo and start typing. The friction is so low that I now use ChatGPT for questions I would have just let pass before.

The desktop app also handles file uploads better. Drag in a PDF and ask questions about it. Take a screenshot and ask for analysis. Paste in a code file and ask for a review. It becomes a natural part of your workflow instead of a destination you have to go visit.

If you're using ChatGPT regularly and haven't installed the desktop app, do it. It's free and makes a bigger difference than you'd expect.

File Upload and Image Analysis Are Underused

ChatGPT can read uploaded files. PDFs, Word documents, spreadsheets, images, code files. Most people I talk to have never tried this.

I upload research papers and ask for summaries. I upload screenshots of error messages and ask what's wrong. I upload meeting transcripts and ask for action items. I upload spreadsheets and ask for analysis.

The quality varies by file type and complexity. Text extraction from images is solid. Understanding the structure of a PDF with tables and charts is less reliable. Code review from uploaded files works well for catching obvious issues.

Is it perfect? Definitely not. Sometimes it misreads charts or misses important context in long documents. But for a first pass or a quick sanity check, it's surprisingly useful. And most people don't even know the feature exists.

When ChatGPT Is the Wrong Tool

There are things ChatGPT is bad at. Knowing when not to use it is as important as knowing how to use it.

Real time information. If you need today's news, stock prices, sports scores, or weather, use a search engine. ChatGPT can access the web if you have Plus and enable search mode, but it's still not its strength.

Sensitive data. Don't paste in your company's financial records, your customers' personal information, or your own passwords and identity documents. The data goes through OpenAI's servers and may be used for training unless you specifically opt out.

Tasks requiring taste and judgment. ChatGPT can suggest ten blog post ideas. It can't tell you which one your audience will actually care about. It can draft an email. It can't tell you whether sending it is a good idea. Use it for drafts and ideas. Use your own judgment for final decisions.

Creative work that needs a unique voice. ChatGPT has a default writing style that's competent but generic. If you need writing that sounds like you specifically, use it for structure and ideas, then rewrite in your own voice. Don't publish raw ChatGPT output as your own writing. People can tell.

FAQ

Q: How do I stop ChatGPT from sounding like a robot?

Tell it explicitly what tone you want. "Write like you're explaining this to a friend at a bar, not like you're writing a corporate memo." Also, give it examples of writing in your preferred style. And iterate. The first draft will still sound somewhat formal. Ask for revisions with specific tone instructions until it matches what you want.

Q: Does ChatGPT learn from my conversations?

It doesn't learn within a conversation beyond what's in the chat history. It doesn't carry knowledge between separate conversations. Each new chat starts fresh. If you want persistent behavior across chats, build a custom GPT with your preferred instructions.

Q: What's the difference between GPT-4o and GPT-4o mini?

GPT-4o is more capable at complex reasoning, creative writing, and nuanced understanding. GPT-4o mini is faster and cheaper but less sophisticated. For most everyday tasks, mini is fine. For work that requires deeper thinking, the full model is noticeably better. Try both and see where you notice the difference.