Prompting

Principles and techniques for writing clear, effective LLM prompts that produce consistent, high-quality output.

Core Principles

Be clear and specific — Treat the model as a skilled worker with zero context. Spell out the task, audience, purpose, and what success looks like. Replace vague quantifiers ("keep it short") with concrete ones ("2-3 sentences").
Say what TO do, not what NOT to do — Positive instructions ("respond in formal tone") outperform negative ones ("don't be casual").
Structure the prompt — Use sections, headers, or delimiters to separate role, instructions, context, examples, and output format.
Set a role — A specific persona improves accuracy, tone, and depth. Be precise: "You are a senior backend engineer reviewing a pull request" beats "You are a developer."
Specify the output format — Never assume defaults. Define: format (bullets, JSON, prose), length, tone, structure.
Provide examples — 3-5 diverse examples dramatically improve output quality. Examples should be relevant, varied, and clearly delimited.
Give context — Who is the audience, what is the purpose, where does this fit in a larger workflow.
Let it think — For complex tasks, instruct step-by-step reasoning. Don't suppress the thinking.
Permit uncertainty — Let the model say "I don't know" rather than fabricate answers.