ChatGPT Prompts: How to Get Better Answers from GPT
How to get sharper, more useful answers from ChatGPT — from custom instructions and roles to iteration and formatting tricks that make a real difference.
Ask ChatGPT "write me a marketing email" and you'll get something that reads like every marketing email ever written. Ask it to write a 90-word email to churned Pilates subscribers, warm but not desperate, ending with a single question, and the output changes completely. The model didn't get smarter between those two requests. You just gave it enough to work with. Most people's disappointing results come down to prompts that could have been typed into any search box a decade ago, and the fix is rarely a longer prompt so much as a more deliberate one.
Hand GPT a role before the task
A role tells the model which slice of its training to lean on. "Explain compound interest" pulls from everything. "You are a high-school math teacher explaining compound interest to a 15-year-old who hates math" narrows the vocabulary, the pacing, and the examples all at once.
Roles work because they set constraints you'd otherwise have to spell out line by line. A "skeptical venture capitalist" will poke holes. A "copy editor who cuts 20% of every draft" will actually cut. Pick the role that behaves the way you need the output to behave.
Notice that the role isn't decoration. It shapes tone (direct but kind), scope (three issues), and depth (one worked fix). If you want a repeatable version of this, save it in your prompt library so you're not rebuilding it every Monday.
Be specific about task, audience, and format
Three variables quietly decide the quality of most answers: what you want done, who it's for, and how it should look on the page. Leave any of them blank and GPT fills the gap with an average guess.
- Task: "summarize" is vague. "Summarize in five bullets, one sentence each, no jargon" is a spec.
- Audience: writing for a CFO and writing for a new intern produce different drafts of the same facts.
- Format: naming the shape up front saves you a round trip. More on that below.
The more your prompt reads like a brief you'd hand a freelancer, the closer the first draft lands. If you want the mechanics of building that brief from scratch, our guide to writing AI prompts walks through it step by step, and the fundamentals piece covers the reasoning behind why structure beats length.
Set Custom Instructions and let memory work
Custom Instructions live in ChatGPT's settings and apply to every chat, so anything true of all your work belongs there rather than in each message. Tell it your role, your tone preferences, the length you like, and what to skip.
Two fields matter. "What would you like ChatGPT to know about you" holds context: your job, your projects, your expertise level. "How would you like ChatGPT to respond" holds behavior: be concise, show code before explanation, skip disclaimers, use British spelling. A useful entry might read: "I'm a product manager at a B2B SaaS company. Default to bullet points. When I ask for copy, give me three options, not one. Don't apologize."
Memory goes a step further, carrying facts across separate conversations so you stop re-explaining that you use Postgres or that your brand voice avoids exclamation marks. You can review and delete what it stores. Treat it like a settings panel you occasionally clean out, not a black box.
Ask for the shape you want
GPT will match a format you request far more reliably than one you hope it guesses. Say the structure out loud.
- Want a comparison? Ask for a table with named columns.
- Want a process? Ask for numbered steps.
- Want options? Ask for exactly three, each with a one-line tradeoff.
- Want something you'll paste elsewhere? Ask for raw Markdown, or JSON with specific keys.
Because you named the columns, you won't get a wandering essay you then have to reformat. When you find a format that works, running it through a prompt optimizer can tighten the wording so the model follows it even more consistently.
Iterate in the same thread
Restarting throws away context the model already has. Instead of opening a fresh chat when a draft misses, correct it in place. "Cut the intro, make the tone drier, and swap the third example for something from healthcare" gets you a targeted revision, not a coin flip.
This is where ChatGPT genuinely shines: it holds the conversation. A good session looks less like a single perfect prompt and more like a short exchange where each turn sharpens the last. Push back the way you would with a collaborator. Tell it what's off, not just that something is.
You can also make the model do the critiquing. Ask it to review its own work before you do:
Self-critique surfaces problems the first pass glossed over, and it costs you one message instead of five rounds of nudging.
Use system-style instructions and step-by-step reasoning
GPT models respond well to firm, system-style direction, which is why the "you are / your job is / always / never" pattern lands so cleanly. Front-load the rules and the model treats them as guardrails for the whole exchange. If you're building something you'll reuse, a dedicated system prompt holds those standing rules so you don't retype them.
For reasoning-heavy work, ask for the thinking, not just the answer. On math, logic puzzles, multi-step planning, or debugging, "work through this step by step and show your reasoning before the final answer" measurably reduces careless mistakes, because the model commits to intermediate steps instead of leaping to a conclusion. On the newer reasoning-focused models this matters less, since they deliberate internally, but a nudge to check its own work rarely hurts.
A few model-specific habits worth keeping:
- For long documents, put your instruction after the pasted text, not before it, so the ask stays fresh.
- When accuracy matters, ask it to flag anything it's unsure about rather than paper over gaps.
- For anything creative, request several distinct directions in one go, then iterate on the one you like.
The same instincts carry to other assistants with small adjustments. If you split your work across tools, the Claude prompting guide covers where the approaches diverge. The core move stays the same everywhere: tell the model who it is, what you need, and what "done" looks like.
