Prompt Engineering Cheatsheet

The core patterns

Pattern	When to use it	Shape
Zero-shot	Simple, well-defined tasks	Just ask.
Few-shot	Anything where the output format is strict	Show 2–5 examples before the real query.
Chain-of-Thought	Reasoning, math, multi-step planning	Ask the model to “think step by step” before answering.
ReAct	Agent / tool-using flows	Interleave Thought → Action → Observation steps.
Self-consistency	High-stakes reasoning where one sample isn’t enough	Sample n completions, take the majority answer.
Reflection	Editing or improving prior output	Ask the model to critique its own draft, then revise.

A reusable template

# Role
You are a <role> helping <audience> with <task>.

# Context
<background the model needs — pasted docs, prior turns, constraints>

# Task
<exactly what you want it to do>

# Output format
- Section 1: ...
- Section 2: ...
- Return as JSON with keys: foo, bar, baz

# Rules
- Do not <thing to avoid>
- If unsure, say "I don't know" instead of guessing.

Anti-patterns (things I stopped doing)

• Stuffing the system prompt with every rule you’ve ever thought of. Keep it < ~400 tokens.
• Using "Please" and "Thank you" as a substitute for clear instructions.
• Asking for JSON without giving a schema example — you’ll get creative keys.
• Re-reading the whole conversation every turn instead of summarizing + pruning.

A quick math bit

Token cost for a prompt of length $n$ tokens run $k$ times with self-consistency:

$$\text{cost} = k \cdot (n \cdot c_{\text{in}} + m \cdot c_{\text{out}})$$

…where $m$ is the average completion length. Useful when deciding if that 10× sampling is worth it.

See also

• LLM model comparison
• RAG architecture diagram