Prompting in 2026: what still matters when the models are this good

In 2023, prompting had spellbooks. People traded magic phrases the way earlier generations traded cheat codes: tell it to think step by step, offer it a tip, inform it that it is a world-class expert, threaten it a little. Some of this actually worked, some of it was folklore, and it was surprisingly hard to tell which was which.

Most of those spells are now dead, and I want to be precise about what killed them, because the conclusion people usually draw ("prompting doesn't matter anymore") is wrong in a way that costs them real results. I've written before about how the whole discipline of instructing AI has climbed a ladder, from wording a prompt, to curating context, to engineering the loop an agent runs in. That post was about the concepts. This one is about the part that is still in your hands when you sit down and type: what, in mid-2026, is actually worth doing.

The spells worked, then the models ate them

The most famous magic phrase had a proper scientific pedigree. In 2022, researchers showed that simply adding "Let's think step by step" to a prompt could lift a model's accuracy on some maths problems from under 20% to nearly 80% (Kojima et al., 2022). That result is one of the most cited findings in the field. For a couple of years, telling the model to reason before answering was the single best trick available.

Then the trick got absorbed into the machinery. Today's reasoning models do that deliberation internally, on their own, before you see a word of output. OpenAI's own guidance for its reasoning models now says plainly: don't prompt them to think step by step, it's unnecessary and can even hurt (OpenAI, n.d.). The best prompt trick of 2022 became a built-in feature, and then briefly an anti-pattern.

Other spells turned out to be mostly noise all along. Researchers who systematically tested prompt variations, including the beloved "I'll tip you" gambit, found the effects small and inconsistent, varying from model to model with no reliable direction (Salinas & Morstatter, 2024). The same research programme found something more unsettling: tiny, meaningless changes, even adding a space at the end of a prompt, could flip a model's answer. A related study measured performance swings of up to 76 percentage points from formatting choices alone in the open models of that generation (Sclar et al., 2024). The honest summary of the spellbook era is that phrasing mattered enormously, but not in ways anyone could actually control.

My favourite illustration of how unstable this ground is: politeness. A 2024 cross-lingual study found that rude prompts degraded performance, and advised against them (Yin et al., 2024). Eighteen months later, a small study on ChatGPT-4o found the opposite, with blunt and even rude prompts slightly outperforming very polite ones (Dobariya & Kumar, 2025). I don't recommend being rude to a chatbot. But tone-level prompt advice has flipped sign within two years, which should permanently lower your confidence in any tip of that species, including the ones in this post.

What still makes the difference

So the models got dramatically better at coping with lazy, vague, badly formatted prompts. Vendors say this almost outright: current guidance from both OpenAI and Anthropic has shifted from clever phrasing towards plain, direct instructions, and Anthropic now frames prompt wording as just one piece of the broader job of managing everything the model sees (Anthropic, 2025). In my own use, the difference is stark. A one-line request that would have produced garbage in 2023 now usually produces something reasonable.

Reasonable, though, is exactly the trap. What I see in workshops, again and again, is people typing a two-line prompt, getting a competent-looking answer, and concluding that this is what the tool can do. The gap between that answer and what the same model produces when it's properly set up is as wide as ever. It has just moved out of the wording and into the setup, which comes down to three unglamorous things.

Context. The model knows nothing about your situation except what you give it. Anthropic's docs suggest treating it like a brilliant new employee who lacks all context on your norms and workflows (Anthropic, n.d.); an earlier version of the same page added "with amnesia," and that's the detail worth keeping, because it forgets your company, your audience, and your standards between every conversation. Paste in the background document. Show an example of what a good version looks like. Say who the output is for and what it will be used for. Every time I watch someone's results jump in a workshop, this is usually the whole explanation: they stopped asking the model to guess things they could have just told it.

Guardrails. Left to itself, a language model optimises for producing a plausible-sounding answer, and it will produce one whether or not the facts cooperate. You have to build the fences yourself, in words: "if you're not sure, say so", "use only the document I gave you", "if information is missing, ask me instead of assuming", "list your sources so I can check them". No wording makes the model honest. The realistic goal is to make its failures visible, and these instructions do that. I've written a separate checklist for catching AI hallucinations; guardrails in the prompt are the front half of that same defence.

Clarity. By this I mean the old, hard discipline of saying what you actually want, which has nothing to do with eloquence or formality. "Make this better" forces the model to guess what better means to you; "cut this to 300 words, keep the argument about costs, lose the anecdotes" doesn't. This is the one part of prompting that was never a trick, and I'd bet on it outliving every model release, because it isn't about models at all. The same discipline applies whenever you hand work to anyone.

And one habit that costs nothing and that almost nobody uses: end your prompt with "ask me clarifying questions before you start." A model that asks a few questions before starting comes back with work that fits the brief far more often. If you take one concrete habit from this post, take that one.

Best before: soon

Now the caveat the genre usually hides in the footnotes, promoted to its own section.

Everything above describes mid-2026, and this field has been demolishing its own advice on roughly an eighteen-month cycle. Step-by-step prompting went from breakthrough to built-in to discouraged, politeness advice inverted between studies, and the tipping trick turned out to be mostly noise. If the pattern holds, some concrete advice in this post will look quaint embarrassingly fast, and I'd rather say that now than have you discover it later.

My guess about what survives: the specific phrasings won't, and the principles will, because the principles aren't really about AI. Give the person doing the work the context they need, tell them the constraints, be clear about what you want, let them ask questions. That was good advice for delegating to humans a century ago. The strange, slightly comic arc of prompt engineering is that we spent three years discovering elaborate tricks for talking to machines, and the tricks are converging on how you'd brief a capable colleague.

So learn the principles and don't get too attached to the tactics. The durable skill sits on either side of the prompt anyway: you still have to decide what is worth asking for, and then judge what comes back.

Sources

Anthropic. (n.d.). Prompting best practices. Claude documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/be-clear-and-direct

Anthropic. (2025). Effective context engineering for AI agents. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

Dobariya, O., & Kumar, A. (2025). Mind your tone: Investigating how prompt politeness affects LLM accuracy. arXiv. https://arxiv.org/abs/2510.04950

Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems. https://arxiv.org/abs/2205.11916

OpenAI. (n.d.). Reasoning best practices. OpenAI API documentation. https://developers.openai.com/api/docs/guides/reasoning-best-practices

Salinas, A., & Morstatter, F. (2024). The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model performance. Findings of the Association for Computational Linguistics: ACL 2024. https://arxiv.org/abs/2401.03729

Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2024). Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. International Conference on Learning Representations. https://arxiv.org/abs/2310.11324

Yin, Z., Wang, H., Horio, K., Kawahara, D., & Sekine, S. (2024). Should we respect LLMs? A cross-lingual study on the influence of prompt politeness on LLM performance. arXiv. https://arxiv.org/abs/2402.14531