OpenAI staff recently shared several tips for getting more out of ChatGPT. I tried them, and it felt like my chatbot got smarter.
The tips came from Christina Kim, a research lead in post-training, and Laurentia Romaniuk, a product manager for model behavior, in an episode of “The OpenAI Podcast” published Wednesday.
1. Ask the hard questions
Kim said users should throw ChatGPT “harder questions” so it can “decide how much it wants to think.” The tougher the prompt, the deeper the reasoning.
Scrolling through my old chats, I realized I’d been doing the reverse — but for good reason. My job is to take complicated ideas and translate them for a general audience, so I mostly use ChatGPT for clarity checks or background research.
This week, I was reading about embodied intelligence and asked ChatGPT an easy question: “What is embodied intelligence?” It gave a clean, simple answer: AI systems embedded in physical agents like robots.
Wanting to understand the mechanics behind the concept, I asked a harder question: “How do robots fuse vision, audio, touch, and feedback in real time?”
That’s when the model shifted gears. It talked about multimodal sensor fusion, specialized encoders, and cross-modal alignment — terms I’d never seen before and sounded like something overheard in a robotics lab.
I asked if this was PhD-level. ChatGPT answered: “Yes — roughly master’s-to-early-PhD level.”
2. Tell ChatGPT who to be
Romaniuk shared that her brother, a biochemical Ph.D., once complained ChatGPT Pro was responding at an undergrad level — until she told him to specify his expertise. After he primed the model as a “frontier researcher,” it produced an insight so advanced it mirrored a breakthrough his lab had made just two weeks earlier.
I’d never experimented with assigning the model a persona, so I figured I’d try it on a low-stakes question: my coffee preference.
As someone who prefers cappuccinos to lattes, my usual explanation is simple: It tastes punchier and less milky.
I asked ChatGPT to become a barista who studied coffee the way sommeliers study wine and explain my preference to me.
The cappuccino-latte divide became a masterclass. It broke the differences into texture, flavor balance, and mouthfeel. A cappuccino is light on top and dense underneath; the foam “lifts” the espresso, sharpening its flavor. A latte, on the other hand, is silkier and more uniform, with milk that folds into the espresso and softens it.
I can now explain my coffee preference with something more informed than “it tastes nicer.”
3. Audit the chatbot’s memory
Romaniuk said memory is one of ChatGPT’s strengths. It lets the model infer what a user really wants or proactively surface information they might care about.
Romaniuk said the way to stay in control is to audit what the model remembers: Delete anything you don’t want it holding onto, or toggle memory on and off in the settings so the chatbot only draws from what you choose.
It’s something I’ve been doing. I don’t turn memory off because most of what I share ends up being useful later. But I do clear out meaningless or throwaway chats so they don’t clutter things or muddy what the model learns about me.
It’s paid off. ChatGPT now understands me as a journalist who interviews AI founders by day and trains for fitness races after hours. It answers like it knows I’m just as likely to ask about model behavior as I am about Hyrox training plans.
4. Ask ChatGPT to improve prompts
Kim said users can ask ChatGPT to help come up with better prompts.
I needed to understand free-electron lasers, which help create the high-energy light sources behind semiconductor manufacturing (for journalism, of course). Instead of pretending I knew what to ask, I asked ChatGPT what questions I should be asking it.
It responded with a set of “high-leverage questions” — grouped from foundational to research-level — that showed me exactly how to think about the technology. For someone without much technical background, this framing was helpful. ChatGPT was teaching me how to ask smarter questions.
5. Switch through personality modes
Romaniuk said she switches between ChatGPT’s personality modes “all the time” to understand how each one feels. It’s part of her job shaping model behavior.
One of her favorites is the “nerd” mode, which gives the model a “very exploratory response style,” she said.
I love a good cynic, so I asked ChatGPT in “cynical mode” to explain embodied intelligence.
It didn’t move my work forward, but the explanation made me laugh: “Embodied intelligence is one of those tech terms people throw around like everyone has a robotics Ph.D. hiding under their bed.”
“People in AI land obsess over this.”
6. Retry tasks regularly and ‘pressure-test’ the model
Romaniuk said she likes to “pressure test” the model — pushing it to its limits to see how it’s changing over time.
For something that couldn’t work now, it might work in three months’ time. “Just keep at it, keep playing, keep trying,” Romaniuk added.
I’ve been “pressure testing” the model through my Korean studies, regularly throwing it prompts that stretch its ability to break down the language. I’ve spent months asking ChatGPT to teach me grammar, extract vocabulary from worksheets, and explain unfamiliar sentence structures.
Earlier versions often pulled out the wrong words or mixed up written forms. Now, it parses text accurately, distinguishes between formal and polite forms, and explains grammar in clear, beginner-friendly steps.
The tips made my use of ChatGPT feel more intentional. Some effects were small, others more striking, but they certainly nudged the model to reveal more of what it can do.
Read the full article here
