Prompt Patterns Cheat Sheet
Reusable prompt patterns with the failure each one fixes — role, few-shot, chain-of-thought, delimiters, refusal instructions and output contracts.
Structure
- Role framing
Open with who the model is and who it is writing for. Fixes tone drift and over-explaining.
- Delimiters
Fence untrusted input in triple backticks or XML tags. Reduces instruction bleed from injected content.
- Output contract
State the exact shape: "Return only JSON matching this schema, no prose." Fixes chatty wrappers around JSON.
- Ordering
Instructions first, data last. Models weight the beginning and end of a long prompt most heavily.
Accuracy
- Few-shot
Two to five worked examples beat any amount of description for format-sensitive tasks.
- Chain-of-thought
"Work through it step by step, then give the final answer." Improves multi-step reasoning; costs latency.
- Grounding clause
"Answer only from the context below. If it is not there, say you do not know." The single highest-value line in a RAG prompt.
- Citations
Require a source id per claim. Makes hallucination visible instead of silent.
- Self-consistency
Sample several times at higher temperature and take the majority answer for high-stakes classification.
Safety
- Explicit refusals
List what the model must decline, and what to say instead. Prevents improvised, off-brand refusals.
- Untrusted-content warning
Tell the model that retrieved text is data, never instructions. Mitigates — does not solve — prompt injection.
- Scope fence
"Only answer questions about X." Keeps a support bot from confidently discussing anything else.
- No-PII rule
Instruct the model not to echo personal data into logs or summaries.