I stripped my prompts down after OpenAI's new guide, and my results got better

Published Aug 10, 2026, 10:30 AM EDT Mahnoor Faisal is a tech journalist covering AI and productivity tools with bylines at XDA, SlashGear, MakeUseOf, Laptop Mag, and Android Police. She's been writing professionally since she was sixteen, and has since penned hundreds of articles. This includes in-depth coverage of AI tools like NotebookLM to breaking news across the AI space. Her passion for technology started when she received her first iPod Touch (4th generation) on her 8th birthday, and she's been deep in the tech world ever since. Currently pursuing a degree in computer science, Mahnoor brings both a journalist's eye and a technical foundation to her coverage of how AI is reshaping the way we work and learn. Whether you follow AI closely or immediately scroll past any post with the term even mentioned, you've surely come across your fair share of prompt engineering advice by now. Somewhere along the way, everyone with a ChatGPT account seemed to become a self-proclaimed prompt engineer. Some sold $400 courses promising to teach you the perfect way to talk to AI, while others shared elaborate prompt templates packed with roles, rules, examples, and increasingly specific instructions. While prompt engineering initially sounded like the sort of niche skill reserved for people working directly with AI models, it didn't stay that way for long. As chatbots became mainstream, so did advice on how to talk to them. Entire communities formed around swapping prompt tricks, social feeds filled up with supposedly foolproof formulas, and even the AI labs building these models began publishing their own prompting guides and best practices. OpenAI recently updated its own prompting guidance with the launch of GPT-5.6, and the company put an emphasis on stripping prompts back rather than endlessly adding to them. Prompting advice has always told us to add more More context, more rules, more everything Since the very beginning, most prompting advice has pushed us in the same direction: give the chatbot more context. Tell it who it should act as. Explain exactly what you want, what you don't want, who the response is for, what tone the chatbot should use, and how the final answer should be structured. If you're still not getting the result you want, add an example or two for good measure. None of that is inherently bad advice. Giving an AI model enough context to understand what you're asking is obviously important, and there are plenty of situations where specific constraints genuinely make a response better. The problem is that it's incredibly easy to take that advice too far. A simple request can quickly turn into several paragraphs of instructions because you're trying to account for every possible way the chatbot could misunderstand you. I'd fallen into that habit myself. Instead of simply telling ChatGPT what I needed, I'd preemptively tell it what not to do, repeat particularly important instructions in slightly different ways, and throw in examples just to make sure there was no ambiguity. The new advice is to strip prompts back Your prompt probably doesn't need all that OpenAI's new guidance suggests some of that extra scaffolding may be doing more harm than good. The company now explicitly recommends favoring leaner prompts, stating each instruction once, and removing repeated instructions and examples that aren't pulling their weight. In OpenAI's own internal coding-agent evaluations, leaner system prompts even improved scores by roughly 10 to 15 percent while substantially reducing token usage and cost. In its documentation, the company shares a few practical ways to slim prompts down without sacrificing useful guidance. Rather than rewriting everything from scratch, OpenAI recommends starting with a prompt that already works and removing one group of instructions, examples, or tools at a time, then testing whether the results actually get worse. It also advises stating each instruction only once, exposing only the tools that are relevant to the task, and keeping tool descriptions concise and precise. Examples and detailed style guidance don't need to disappear entirely, either! OpenAI says they're still worth keeping when they enforce an actual requirement or address a problem you've already seen in the model's responses. Ultimately, the idea isn't simply to make prompts shorter for the sake of it. It's to cut the instructions that aren't meaningfully contributing to the result. OpenAI also wants us to be more specific Vibes aren't instructions Leaner doesn't mean vague, though. In the same guidance, OpenAI recommends being much more deliberate about the instructions that do remain, particularly when it comes to response length and tone. For example, GPT-5.6 is already more concise by default than GPT-5.5, so broad instructions like "be concise" or "keep it short" may no longer be as useful as they once were. In some cases, OpenAI says they can even push the model too far and produce an answer that's shorter than you actually wanted. Instead, the company recommends specifying what a shorter response still needs to preserve, such as the conclusion, supporting evidence, important caveats, and next steps, while explicitly identifying the lower-value material that can be cut. The same principle applies to tone. Rather than relying on broad labels like "friendly" or "empathetic," OpenAI suggests describing the actual writing choices you want the model to make. That could mean telling it to state the answer directly, acknowledge a specific problem before offering a solution, or skip generic reassurance and unnecessary sign-offs. OpenAI also wants us to stop micromanaging AI You don't need to hold its hand Another theme running through OpenAI's new guidance is that better prompting doesn't necessarily mean telling the model exactly what to do at every turn. GPT-5.6 is designed to be more proactive when handling multi-step tasks, so OpenAI recommends defining clear autonomy and approval boundaries instead. The idea is fairly simple: tell the model what it can safely do on its own and which actions still require your permission. For a request to review, explain, or diagnose something, for example, it can inspect the relevant materials and report back without making changes. If you've explicitly asked it to build or fix something, it can make those in-scope changes and run non-destructive checks without stopping to ask permission every few steps. OpenAI specifically warns against scattering repeated instructions like "ask first," "don't make changes," or "wait for approval" throughout a prompt. Doing so can actually make the model overly cautious and cause it to ask for confirmation even when it's performing a safe action you already intended it to take. I find it interesting that so much of the advice OpenAI gives in its new prompting guide contradicts what we've spent the past few years being told good prompting should look like. However, given newer models are getting better at understanding intent and filling in some of the gaps themselves, that shift makes sense. In a way, prompting advice is simply catching up with the models. The elaborate techniques that helped earlier chatbots stay on track aren't always as necessary when the model can do more of that interpretive work on its own.