Prompt engineering changed faster than most fields. Many 2023-era 'tricks' (let's think step by step, role play as expert) became unnecessary as models got better. Some structural patterns remain durably useful.
What still matters
Clear task description. Few-shot examples for unusual formats. Output schema (JSON structure). Explicit constraints ('do not include X'). These remain high-leverage and increasingly the foundation for the model interpreting intent.
What faded
'Let's think step by step' — modern models reason by default; this rarely helps. Role-play preambles ('you are an expert X') — minimal effect on factual tasks. Threats / bribes — measurably negative on recent models.
What grew
Structured input (clear delimiters, labeled sections). Tool descriptions (the model picks tools from descriptions). Context management (knowing what to include and what to summarize). All these are now the bulk of prompt-engineering effort.