Core prompting patterns

Clear, specific instructions with context, examples and an explicit output format beat clever tricks.

Stable: this rarely changes. Reviewed 17 Oct 2026.

Key ideas

• Give the model a role, the goal, who the output is for and what good looks like.
• Separate instructions from data with XML tags or delimiters, especially for untrusted input.
• Few-shot: 2 to 5 diverse examples, including edge cases. Examples steer format strongly.
• Ask for reasoning before the answer on complex tasks, or use a reasoning model.
• Say what to do rather than only what not to do, and explain why a rule exists.
• Version prompts in code, review them like code, and run evals on every change.

Common mistakes

• Long lists of ALL-CAPS rules. Modern models over-apply them.
• Changing prompts in production without a regression eval.

Go deeper

• Prompt engineering overview (Claude docs)