The handful of prompting moves that actually matter.
Most prompting advice is folk medicine. A small number of moves consistently work; the rest is noise. These are the ones worth keeping.
What works
Be explicit about the task. "Summarize this in 3 bullets" beats "summarize this." Specify length, format, audience, what to skip.
Show, don't tell. 2-3 input/output examples (few-shot) shape the model better than a paragraph of rules.
Ask for structured output. JSON with a schema. Or use the provider's structured output API. Free-text output is harder to use downstream.
Put instructions where the model attends. System prompt up top, user message at the bottom. Repeat critical constraints right before the answer.
Let it think before answering. For reasoning tasks, add "think step by step" or use a thinking model. For simple tasks, skip it.
Use XML or Markdown to delimit sections. Anthropic models love XML tags. OpenAI is fine with Markdown headings.
Structure with XML tags
For Claude in particular, wrapping inputs in semantic tags consistently outperforms inline prose instructions.
<task>Classify the support ticket below as one of: billing, technical, account, other.</task><ticket>{ticket_body}</ticket>Return only the label, no other text.
What stopped working
Role-play prompts. "You are a world-class expert..." adds nothing on modern models.
Threats and rewards. "I'll tip you $200" was a meme that briefly worked and then didn't.
Hedging language. "If possible, please try to..." invites the model to skip the task.
Long preambles. Get to the task fast. Save tokens.