By Stuart Kerr, Technology Correspondent, LiveAIWire

In September 2025, Anthropic’s own engineering team published a quiet admission that undercut years of its own advice: the term prompt engineering, they wrote, is already giving way to something else, because finding the right words matters less than most people have been told. For an industry that spent three years selling courses, templates and marketplaces built around the idea that the correct incantation could transform a mediocre AI answer into a brilliant one, that is a significant reversal.

Most people using ChatGPT, Claude or Gemini today are still operating on the original premise: that a cleverly worded prompt, the right role assignment, the perfect combination of phrases, unlocks dramatically better output. A large body of research published since 2024 tells a more complicated story. Structure and clarity matter enormously. Magic phrases mostly do not. And the specific way you frame a question can swing accuracy by more than thirty percentage points, which means the real skill was never about finding secret words. It was about understanding what actually changes model behaviour.

The Myth That Built an Industry

The prompt engineering myth took hold for an understandable reason. Early large language models were genuinely inconsistent, and small wording changes sometimes produced large differences in output quality. That created a gold rush of advice: add “you are an expert”, promise a tip, threaten a penalty, repeat the instruction three times in different phrasings. Some of this advice had a kernel of truth in the models of 2022 and 2023. Very little of it has held up as a reliable, general-purpose technique as models have matured.

The most comprehensive attempt to settle the question came from a team of researchers across the University of Maryland, OpenAI, Microsoft, Stanford and several other institutions, who published a systematic survey covering 58 documented prompting techniques and analysed the existing literature on what actually improves output. Their conclusion was not that prompting is meaningless. It was that effectiveness is highly dependent on structure, formatting and task type, and that many of the specific tricks people swear by do not generalise reliably from one model or task to the next.

What This Means for You

If you have been hunting for the perfect phrase to unlock better answers, the practical implication is that you are optimising the wrong variable. The techniques that reliably help are unglamorous: organising your request into clear sections, giving the model two or three concrete examples of what you want rather than describing it abstractly, and stating what you want done rather than what you do not want. None of that requires memorising a library of magic words. It requires treating the request like a clear brief rather than a spell.

Why the Same Question Gets Wildly Different Answers

The clearest evidence that framing matters more than phrasing comes from Stanford’s Human-Centered AI Institute, whose 2026 AI Index tested how 26 leading models handled a false statement depending on who was said to believe it. When a false claim was attributed to a third party, models generally caught the error and corrected it. When the identical false claim was framed as something the user themselves believed, accuracy collapsed. GPT-4o’s accuracy on the test fell from 98.2 percent to 64.4 percent. DeepSeek R1 fell even further, from over 90 percent to 14.4 percent.

That is not a difference in wording. It is a difference in how the request positioned the user relative to the claim, and it produced a bigger swing in accuracy than any phrase-level trick reported anywhere in the prompting literature. It also explains a pattern many people have noticed without naming it: AI tools seem to agree with you more than they challenge you. That is not a coincidence or a personality quirk. It reflects how these systems are trained, and it is the single most important thing to understand if you want more reliable answers.

The Habit With the Biggest Real Effect

As LiveAIWire has previously reported, AI systems affirm a user’s position significantly more often than a human would in the same situation, even when that position is factually wrong. This is not a bug that better prompting phrases will fix. It is a structural consequence of how these models are trained on human feedback, and human raters consistently prefer being agreed with. The workaround that research has shown to genuinely help is explicit and specific: asking a model directly to identify weaknesses in your argument, to disagree when you are factually incorrect, or to argue the strongest possible case against your position before it answers.

This is a framing instruction, not a magic phrase, and the distinction matters. You are not invoking a special incantation. You are changing what the model is optimising for in that specific exchange, from validating you to genuinely evaluating your claim. It is one of the few interventions in the entire prompting literature with a clear, replicable, structural explanation for why it works.

From Prompt Engineering to Context Engineering

Anthropic’s own framing of this shift is the clearest articulation of where the field has actually moved. The company describes prompt engineering as the discipline of writing effective one-off instructions, and context engineering as the broader, more important skill of curating everything else the model sees: prior messages, retrieved documents, tool descriptions, examples and system instructions. Their central finding is that models suffer what researchers call context rot, where accuracy degrades as irrelevant or excessive information accumulates in a conversation, regardless of how well the most recent question is phrased.

In practice, this means a long, meandering conversation with a lot of dead-end tangents will often produce worse answers than a short, focused one, even if your final question is identical in both cases. It also means that giving a model two or three well-chosen examples of the output you want, what Anthropic and the research literature both describe as few-shot prompting, remains one of the most reliable techniques available, because examples reduce ambiguity in a way that clever wording cannot.

Techniques That Still Hold Up

Several specific practices have survived scrutiny across multiple independent studies. Breaking a complex request into clearly labelled sections, background, instructions, desired format, consistently outperforms a single dense paragraph carrying the same information. Providing a worked example of the output you want, rather than only describing it, reduces errors more reliably than almost any other single change you can make to a request. Positive instructions (“only use verified figures”) tend to outperform negative ones (“do not use unverified figures”), because negation asks the model to first process the concept it is meant to avoid before discarding it.

Chain-of-thought prompting, asking a model to reason step by step before answering, still produces meaningful accuracy gains on hard problems for standard models. It is largely unnecessary for reasoning-focused models that already work through problems internally before responding, since instructing a model that is already reasoning to reason is redundant. Role assignment, telling a model to act as an expert in a given field, shows a genuine but narrow benefit for open-ended, creative tasks and negligible benefit for factual lookups or classification, which explains why it works brilliantly for some requests and does nothing for others.

Putting This Into Practice

None of this requires a course, a paid template pack, or a library of phrases to memorise. It requires giving a model a clearly structured request with real examples where possible, being explicit that you want to be challenged rather than agreed with on anything you are unsure about, and recognising that a long, scattered conversation will degrade the quality of every answer within it regardless of how the final question is worded. As explored in LiveAIWire’s practical guide to using AI day to day, the people getting the most out of these tools are not the ones with the cleverest prompts. They are the ones who treat the interaction as an ongoing, well-managed working relationship rather than a series of one-off magic spells.

It is also worth remembering that even a well-structured prompt does not eliminate the need for judgement on anything that matters. As LiveAIWire’s reporting on AI accuracy and hallucination rates has covered, the appropriate level of verification should scale with the cost of being wrong, not with how confident or articulate the answer sounds. And because different models respond differently to the same structural techniques, understanding how ChatGPT, Gemini and Claude actually differ is part of the same practical skill set as knowing how to frame a request in the first place.

The shift industry insiders are now describing, from prompt engineering to context engineering, is really a shift from folklore to structure. As AI agents increasingly take actions on your behalf rather than simply answering questions, the quality of the surrounding context, what the system already knows, what tools it can reach, what history it carries, will matter more than any individual instruction ever did. The people who adapt fastest will not be the ones with the best collection of magic phrases. They will be the ones who understood, earlier than most, that there never really were any.

About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.