Prompt Engineering Is Dead. Here's What Actually Works.
Modern models understand intent. The best prompts describe the goal and constraints, not a series of instructions.
What Prompt Engineering Used to Mean
In 2022 and 2023, prompt engineering meant learning specific syntax that would trigger better outputs from models that were less capable at inference. "Act as a [role]", "Think step by step", "Let's approach this from the perspective of..." — these were genuine improvements on earlier models because the models needed to be guided explicitly. The advice was accurate for its time.
Why Most of It Is Obsolete
Claude Opus, GPT-4, and Gemini Ultra understand context and intent well enough that much of the explicit guidance is unnecessary. Bottley ran a controlled comparison: 50 tasks each with traditional "optimized" prompts versus direct, conversational descriptions of the goal. The direct descriptions matched or outperformed the optimized prompts in 41 of 50 cases. The remaining 9 cases involved highly structured output formats where explicit templates still helped.
What Actually Works Now
Three things produce better output reliably: specificity about the end state ("I need a 300-word product description that emphasizes durability, for an audience of outdoor enthusiasts, in a direct tone without superlatives"), examples of the format or tone you want (one sentence of what good looks like), and constraints on what to avoid. These are not prompt engineering tricks — they are good communication.
The One Technique That Still Matters
The only prompt engineering approach that consistently improves output on modern models is providing a brief example of what good looks like — one or two sentences in the style you want. Models respond to demonstrated style more reliably than to described style. "Write this like the opening of The New Yorker" produces different output than "Write this in a sophisticated, literary style." The example is the instruction.