Prompt Engineering Cheatsheet

Structure every prompt

1. Role — who the model should act as ("You are a meticulous reviewer").
2. Task — the specific job, in one sentence.
3. Context — the inputs, in [Brackets] so they're fillable.
4. Format — exactly how to shape the output (sections, table, JSON).

Patterns that work

• Output contract — "Return exactly these sections: …" beats "please summarize nicely."
• Few-shot — one good example teaches a tricky format faster than three paragraphs of rules.
• Critique-then-fix — ask for problems first, then the corrected version; better than "improve this."
• Escape hatch — "If you're not sure, say what's missing" prevents confident nonsense.

Fixing a prompt that misbehaves

• Output too vague → constrain length, tone, and format explicitly.
• Ignores an instruction → move it to the end, or make it a numbered requirement.
• Invents details → provide the facts in context and say "only use what's given."
• Inconsistent between runs → shrink the ask; one task per prompt.