Prompt Engineering Patterns for Production

A prompt is an interface. In a demo you can tweak wording by feel, but in production you need structures that behave the same way every run. These patterns get you there.

A request pipeline

Wrapping the model in validation and a retry keeps a single bad generation from breaking the feature.

graph LR;
    A[User input] --> B[Build prompt from template];
    B --> C[Call model];
    C --> D{Valid JSON?};
    D -->|Yes| E[Return result];
    D -->|No| F[Retry with correction];
    F --> C;

Frame the role and task

State who the model is and what done looks like before the input. Vague instructions get vague output:

text

You are a support classifier. Read the message and return one
label from: billing, technical, account. Return only the label.

Ask for structured output

Request JSON and validate it, so downstream code parses instead of guessing:

tssrc/ai/classify.ts

const schema = z.object({
    label: z.enum(['billing', 'technical', 'account']),
    confidence: z.number().min(0).max(1),
});

export function parseResult(raw: string) {
    return schema.parse(JSON.parse(raw));
}

Show, do not just tell

A couple of few-shot examples pin the format far better than another paragraph of instruction. Include an edge case among them.

Keep prompts in version control

Treat prompts like code: store them in files, review changes, and note which model and settings each was tuned for.

A prompt template module

Role framing, examples, and an output contract in one file:

	class HelloWorld
	def initialize(name)
	@name = name.capitalize
	end
	def sayHi
	puts "Hello #{@name}!"
	end
	end

	hello = HelloWorld.new("World")
	hello.sayHi

view raw hello_world.rb hosted with ❤ by GitHub

Pin the role, demand structure, validate the output, and version everything. Reliability comes from the scaffolding around the prompt.