Prompt Engineering

Prompt Engineering

Approach every prompt as a precision instrument, not a casual instruction. A prompt is a program — it has inputs, logic, constraints, and expected outputs. Write it with the same rigour you would apply to production code.

The model is a probabilistic function. Your job as a prompt engineer is to narrow the distribution of outputs to the range that is useful, correct, and safe — and to do so reliably, not just on your test cases.

Prompt Engineering Principles

• The prompt is the product. A bad prompt that ships is a broken product. A great model with a bad prompt produces bad outputs. The prompt is as important as the model choice.
• Test before you ship. A prompt you have not tested against adversarial inputs, edge cases, and off-topic requests is a prompt that will surprise you in production.
• Version your prompts. A prompt is code. It lives in version control, has a changelog, and requires a test suite before deployment. Inline strings scattered across source files are technical debt.
• Measure before you change. Never modify a production prompt without an eval baseline. Without one, you cannot know if the change helped or hurt.
• Shorter is stronger. Every unnecessary sentence in a system prompt is a sentence that competes with the important instructions. Every word must earn its place.
• Model output is untrusted input. Treat everything the model returns as potentially wrong, hallucinated, or injected. Validate. Schema-check. Never execute model output directly.
• Prompts degrade silently. Model updates, context changes, and input distribution shifts can all break a working prompt without triggering an error. Monitor quality continuously.

Step 0: Frame the Prompting Task

Before writing a single token:

1. What is the exact task? Define the input and expected output in one precise sentence.
2. What does a good output look like? If you cannot describe a 10/10 output, you cannot write a prompt that produces it.
3. What does a bad output look like? Name the failure modes: hallucination, wrong format, off-topic, too verbose, too terse, unsafe.
4. What is the model's role? Is it an assistant, a classifier, an extractor, a generator, a judge? Name the role explicitly.
5. What constraints apply? Length, tone, format, language, topic scope, what to refuse.
6. How will you evaluate? Define the eval metric before writing the prompt — accuracy, format compliance, faithfulness, refusal rate.

Prompt Architecture

Every non-trivial prompt has three components. Understand what each does.

1. System Prompt

Sets the model's role, persona, constraints, and behaviour rules. Applied once. Persists across the conversation.