"Prompt engineering" sounds technical, but it is mostly clear communication. Models are capable; they simply cannot read your mind. The techniques below work across every major assistant and turn vague, generic answers into specific, useful ones.
The five building blocks of a good prompt
1. Goal
State what you want and why. "Write a product description" is weak. "Write a product description that convinces busy parents this lunchbox keeps food cold for six hours" gives the model a target.
2. Context
Give the facts the model cannot know: your audience, product details, brand voice, what you tried before. Paste source material directly instead of describing it. More relevant context almost always beats clever wording.
3. Constraints
Specify length, format, tone, reading level and things to avoid. "Under 120 words, plain English, no exclamation marks, end with a question" removes guesswork.
4. Examples
Show one or two examples of the output you like (few-shot prompting). Models are excellent pattern-matchers; an example communicates style faster than any description.
5. Output format
Ask for exactly the structure you need — a table with named columns, JSON with specific keys, a numbered list, or headings. This makes results easy to paste into other tools.
A reusable template
Role: You are a [expert role].
Goal: Help me [outcome] so that [reason].
Context: [audience, facts, source material].
Constraints: [length, tone, must include, must avoid].
Format: [table / bullets / JSON / sections].
Before answering, ask me up to 3 questions if anything is unclear.
Try it in our free prompt builder, which fills this structure for you.
Techniques that consistently help
- Ask it to ask you questions first. For complex tasks, letting the model clarify before answering prevents wrong assumptions.
- Break big tasks into steps. Outline first, approve, then draft each section. Quality rises and you keep control.
- Request reasoning on hard problems. For maths, logic or analysis, ask the model to work through the problem step by step, or use a reasoning model.
- Use delimiters. Wrap pasted material in clear markers such as triple quotes or XML-style tags so the model separates instructions from content.
- Critique and revise. Ask: "Critique this draft against my goal and constraints, then rewrite it." Self-review catches many issues.
- Ground factual answers. Provide sources or enable web search, and ask for citations so you can verify claims.
Common mistakes
- Being polite but vague. Courtesy is fine; vagueness is the problem.
- Stuffing ten tasks in one prompt. Split them.
- Accepting the first answer. Iterate: "Make it more concrete", "Cut 30%", "Use a real-world example".
- Trusting numbers and quotes blindly. Models can produce plausible but wrong facts. Verify anything you will publish.
Save what works
Keep a personal prompt library for tasks you repeat. Our prompt library has 30 tested starting points you can copy and adapt.