LLM Prompting Cheatsheet

Patterns that actually move model performance — not magic words.

1 credit

Structure

  • **Role**: start with what it is ("You are a code reviewer") — anchors tone.
  • **Goal**: state the outcome in one sentence, not the steps.
  • **Context**: relevant facts — project, constraints, data samples — top of prompt.
  • **Format**: explicit output shape. "Return JSON with keys: a, b, c. No prose."
  • **Examples (few-shot)**: 1-3 input→output pairs. This beats clever instructions.

Techniques that help

5 items
Chain-of-thought
"Think step-by-step" before the answer — improves multi-step reasoning
Self-critique
"Draft, then critique, then rewrite" — catches obvious errors
Decomposition
Break a big task into sequential prompts — each prompt's output feeds next
Grounding
Paste relevant docs/code in context instead of assuming memorized knowledge
Constraints
"Max 100 words" / "Use only listed functions" — limits hallucination

Anti-patterns (don't bother)

  • "You are an expert" stacked 5 ways — marginal at best.
  • Threats or flattery — modern models ignore these.
  • Asking for sources without grounding — it'll invent URLs.
  • Super-long instructions + tiny context — it'll skip instructions mid-completion.

Debugging bad output

  • Lower temperature if output is random / inconsistent.
  • Move important constraints to the **end** of the prompt (recency bias).
  • Ask it to **echo its understanding** first — catches misread instructions.
  • Test with a smaller, clearly-wrong example — if it still fails, the prompt is broken.

Further reading