OpenAI staff shared six practical tips for squeezing more value from ChatGPT, and testing them made the chatbot feel noticeably smarter.
The guidance came from Christina Kim, a research lead in post-training, and Laurentia Romaniuk, a product manager focused on model behavior, during a recent episode of The OpenAI Podcast. Here’s how each tip played out when put into practice.
- Ask the hard questions Kim suggests pushing ChatGPT with tougher prompts so the model can show its reasoning more fully. In my own experience, I’d often start with simpler queries to quickly surface clarity or do background checks for journalism. Curious about embodied intelligence, I first asked a straightforward question: “What is embodied intelligence?” The answer was clean and concise, describing AI systems integrated into physical agents like robots.
To dive deeper, I followed up with a more challenging question: “How do robots fuse vision, audio, touch, and feedback in real time?” That probe triggered a shift in the model’s response, introducing terms such as multimodal sensor fusion, specialized encoders, and cross-modal alignment—concepts I’d rarely seen outside robotics literature. When I asked whether this was PhD-level material, the model replied that it was roughly at the master’s-to-early-PhD level.
- Tell ChatGPT who to be Romaniuk shared a telling example: her brother, a biochemical Ph.D., complained that ChatGPT Pro answered at an undergraduate level until she instructed the model to assume the role of a “frontier researcher.” With that priming, the model produced insights so advanced they mirrored a breakthrough from his lab two weeks earlier.
I tried this with a low-stakes scenario of my own: explaining my coffee preference. I, a cappuccino fan, asked ChatGPT to adopt the persona of a barista who studies coffee as rigorously as a sommelier studies wine. The result was a detailed breakdown of the cappuccino–latte distinction in terms of texture, flavor balance, and mouthfeel. A cappuccino’s foam sits atop and lifts the espresso, sharpening its flavor, while a latte yields a silkier, more uniform result as milk blends with the espresso. Now I can articulate my preference with a more informed, evidence-based explanation.
- Audit the chatbot’s memory Memory is one of ChatGPT’s strengths, enabling the model to infer needs and surface relevant information proactively. Romaniuk emphasizes maintaining control by auditing what the model remembers: delete anything you don’t want retained, or toggle memory on and off so the chatbot draws only from chosen data.
Personally, I don’t keep memory turned off because most shared information proves useful later. Yet I do regularly prune meaningless chats so they don’t clutter the memory or misguide the model about my preferences. The payoff is tangible: ChatGPT now recognizes me as a journalist who interviews AI founders and as someone who trains for fitness races, so it answers with an awareness of both interests.
Ask ChatGPT to improve prompts
Kim advises asking ChatGPT to help craft better prompts. I needed background on free-electron lasers, which power high-energy light sources used in semiconductor manufacturing. Rather than guess the right questions, I asked ChatGPT to propose a set of high-impact questions, organized from foundational to research-level. The suggestions helped me frame my inquiry and prompted smarter lines of inquiry, especially helpful for non-technical readers who benefit from clear, guided thinking.Switch through personality modes
Romaniuk routinely cycles through ChatGPT’s personality modes to gauge how each one feels, which is useful for shaping model behavior. One favorite is the “nerd” mode, described as delivering a very exploratory response style.
I tried a playful experiment by asking for a cynical mode to explain embodied intelligence. The result didn’t propel my work forward, but it was entertaining: the model quipped that embodied intelligence is one of those tech terms people toss around as if everyone has a robotics Ph.D. hidden under the bed. It’s a reminder of how different tones can illuminate or entertain, even if not always advancing the task.
- Retry tasks regularly and pressure-test the model Romaniuk advocates pressure-testing the model—pushing it to its limits to observe how it evolves over time. Something that seems unlikely today may become feasible in a few months with continued experimentation. Her advice: keep practicing, keep testing.
That approach has also informed my own study of Korean. I regularly challenge ChatGPT with prompts to dissect grammar, extract vocabulary from worksheets, and explain unfamiliar sentence structures. Early model versions sometimes mishandled word choices or written forms, but improvements over time have led to accurate parsing, clear explanations of formal versus informal language, and beginner-friendly grammar guidance.
Overall, these six tips didn’t just make ChatGPT feel more capable—they encouraged a more intentional, exploratory way of using the tool. Some effects were subtle, others more pronounced, but the result is a more powerful assistant that can reveal capabilities beyond initial impressions.
If you’re curious to try these ideas, consider how you might tailor them to your own workflow or field. Which tip resonates most with your work, and what results do you hope to achieve by pushing the model a bit further each day?
