A lot of prompt engineering advice still circulating online is built for a generation of AI models that no longer exists. Reasoning-capable models changed the rules enough that OpenAI and Anthropic have both published official guidance walking back techniques they used to recommend. Here is what current, primary-sourced guidance actually says.
Stop telling reasoning models to "think step by step"
This is the single most outdated piece of advice still in wide circulation. OpenAI's official documentation for its reasoning models states directly: "prompting them to think step by step or explain your reasoning is unnecessary," because these models already reason internally before answering. Anthropic's current guidance for Claude points the same direction, its newest models use what Anthropic calls adaptive thinking, where the model itself decides how much internal reasoning a question needs, and Anthropic explicitly recommends a high-level instruction like "think carefully about this" over prescriptive step-by-step scaffolding. For both companies, this is a genuine reversal from older guidance, not a minor tweak.
Shorter prompts measurably beat longer ones
OpenAI's own internal testing, published in its GPT-5 prompting guide, found that leaner, single-instruction prompts scored 10 to 15% higher than prompts that restated the same instruction multiple times for emphasis, while using up to 66% fewer tokens. State each instruction once. Repeating it for reinforcement does not help and measurably hurts.
Stop yelling at the model
Anthropic's official guidance contains a specific, testable finding: phrases like "CRITICAL!", "YOU MUST", and "NEVER EVER" produce worse results on current Claude models than calm, direct instructions. Positive framing also outperforms negation consistently, "only use real data" works better than "don't use mock data." This is a real behavioral finding from the model's own maker, not a stylistic opinion.
The research finding that undercuts a popular technique
This is genuinely surprising and worth knowing before you build a habit around it. A 2026 study tested adding expert personas to system prompts, telling a model "you are a world-class doctor" or similar, across 2,410 factual questions and four model families. Persona prompting did not improve factual accuracy. On the MMLU benchmark specifically, accuracy actually dropped, from 71.6% with no persona to 68.0% with a short persona, and down to 66.3% with a longer, more elaborate one. Persona prompting still has a place, it can help with tone and style for creative or open-ended writing, but for anything where factual accuracy matters, it appears to actively hurt rather than help.
Where vendors actually disagree
Not everything is settled. Google's official Gemini prompting documentation still recommends including few-shot examples by default, stating that prompts without them "are likely to be less effective." OpenAI and Anthropic's current guidance, by contrast, recommends starting with a zero-shot prompt and only adding examples if testing shows they help. This is a real, current disagreement between major AI labs, not a settled best practice, worth knowing so you do not treat either position as universal.
What still matters
Across all three vendors, a few things hold steady: be specific about the exact output you want rather than the general topic, state the format explicitly if you want a table or a specific list length, give the model context it cannot otherwise guess, and iterate by telling the model exactly what to change rather than rewriting your prompt from scratch. Good prompting increasingly looks less like a technical skill and more like clearly delegating a task to a capable colleague.
The takeaway
If you learned to prompt AI more than a year or two ago, some of what you learned is now actively counterproductive on current reasoning-capable models. The single highest-value update: stop adding "think step by step," and stop assigning expert personas when accuracy is what actually matters.
