prompt-engineering

Start with a direct instruction, add constraints only when needed, and include 2-5 input-output examples for consistent formatting. Match the level of specificity to the task's fragility, using exact scripts for critical operations and open guidance for flexible tasks.

Few-shot prompting teaches the model by showing 2-5 input-output example pairs instead of explaining rules. Use it when you need consistent formatting, specific reasoning patterns, or edge-case handling, balancing example count against token cost.

Remove redundant words, move stable instructions into system prompts, consolidate similar rules, and assume the model already knows common concepts. Challenge each paragraph by asking whether it justifies its token cost.

Use chain-of-thought prompting for complex problems requiring multi-step logic, mathematical reasoning, or verifiable thought processes. Adding step-by-step reasoning improves accuracy on analytical tasks by 30-50%.

Vague or optional phrasing lets the model rationalize skipping steps. Use imperative language like "YOU MUST", require explicit announcements of actions, and add verification steps so compliance becomes automatic rather than discretionary.

Persuasion principles like authority and commitment improve compliance but should not be combined excessively or used manipulatively. Avoid liking and reciprocity for discipline enforcement, as they create sycophancy and conflict with honest feedback.