Prompt Engineering
Beginner 11 min read Prompt Engineering

Prompting Basics

The cheapest lever in AI engineering

Prompts are the primary interface to an LLM: role, instructions, context, output format and constraints. Well-structured prompts get dramatically better results with zero infrastructure.

promptinstructionsoutput formatconstraintsroleprompt structure
Prerequisites: System vs User Prompts
Before we start

What you will learn

  • Structure prompts role → task → constraints → format
  • Write stable, testable prompt text
  • Iterate on prompts like code
Introduction

What is this?

A prompt is a set of instructions + context handed to an LLM. Its anatomy: role (who the model plays), task (what to do), context (what to use), constraints (what not to do), output format (how the answer should look). Ordering and clarity measurably change quality — prompting is engineering, not magical phrasing.

Motivation

Why do you need it?

  • It's the only lever that costs zero infra and works on every model.
  • Clear constraints prevent the 'fluent-but-off' failure mode.
  • Structured prompts are debuggable and A/B-testable.
  • Small refinements compound across thousands of calls.
Think about it

Mental model

Imagine

A clear brief beats a vague one

Think of

Prompting

Two freelancers get the same job: one is told 'summarize it', the other 'you're a reviewer; summarize in 3 bullet points, 40 words max, flag anything unverified'. The second brief produces usable work the first time — prompting is writing that brief.

Mechanics

How it works

  • Set the role: 'You are a concise support specialist.'
  • State the task in one imperative sentence.
  • Include the data/context the model should use.
  • Add constraints: length, tone, banned claims, required fields.
  • Specify output format (bullet, JSON, table) and re-state key rules at the end.

A stable prompt skeleton.

Role
Task
Context
Constraints
Format
Real world

In production

  • System prompts follow the same skeleton and stay versioned.
  • Prompt files live in the repo, diffable like code.
  • Template variables (user data) fill context slots safely.
Guidelines

Best practices

  • Keep prompts concrete: numbers, examples, explicit negatives.
  • End with a one-line restatement of the key constraint.
  • Version prompts; ship changes through the same review pipeline.
  • Prefer explicit rules over 'be careful' prose.
Watch out

Common mistakes

  • One giant prose paragraph instead of structured instructions.
  • Over-prompting ('be super ultra helpful') that dilutes real rules.
  • Editing prompts in production dashboards without logs/rollback.
Root cause analysis

Common issues

Error
Outputs ignore the instruction entirely
Root cause

Prompt buried under context or ambiguous role

Fix

Restructure role/task before data; re-state rules at the end.

Error
Strange regressions with no prompt change
Root cause

Model/provider offset drift

Fix

Pin model version and re-run prompt regression tests.

When things break

Debugging tips

  • Output drifts from format → the format instruction is vague; give a schema.
  • Model ignores a constraint → move it to the start AND end of the prompt.
  • Behavior changed after a model bump → re-check prompt anatomy against the new model.
Interviews

Interview checkpoint

Test yourself

Quick quiz

0/2
1

Which element is LEAST useful in a prompt?

2

Key constraints should be restated near the end of a prompt.

Go deeper

Official documentation

Prompt engineering guide (OpenAI)
docs
Prompt engineering (Anthropic)
docs