prompt engineering cheatsheet

Prompt engineering is less about magic words and more about giving the model the right context, constraints, and output format. These are the patterns builders in the network use daily.

system prompt structure

A reliable system prompt has four parts:

1. Role — who the model is (you are a code reviewer focused on security)
2. Context — what it knows (the user is building a Next.js app with Supabase)
3. Constraints — what it must and must not do (never suggest deprecated APIs)
4. Output format — how to respond (respond in JSON with keys: issue, severity, fix)

Keep it under 500 tokens when possible. Long system prompts dilute attention.

few-shot beats instructions

One good example is worth ten rules. Show the model what you want:

Input: "refactor this to use async/await"
Output: [show the exact refactor style you expect]

Two to three examples cover most formatting and tone requirements. More than five usually hurts generalization.

structured outputs

For production, always request structured output when you need to parse results:

• Use JSON mode or response schemas where supported
• Define required fields explicitly
• Include an error or unable_to_complete field for graceful failure

Parsing free-form text in production is a bug waiting to happen.

chain of thought (when it helps)

Ask the model to reason step-by-step for:

• Multi-step math or logic
• Complex code debugging
• Decisions with trade-offs

Skip it for simple transformations — it adds latency and cost without benefit.

testing prompts

Treat prompts like code:

• Keep a test set of 10–20 real inputs
• Run them on every prompt change
• Score outputs manually or with a second model pass

The builders who ship reliable AI features version their prompts in git and regression-test them.

quick reference

Pattern	Use when
System role + constraints	Every production prompt
Few-shot examples	Formatting, tone, edge cases
JSON / schema output	Anything downstream parses
Step-by-step reasoning	Complex logic, debugging
Delimiters (""", ###)	Separating user content from instructions

Download the full cheatsheet below for copy-paste templates.