How do you write better AI prompts?

A prompt is just your instruction to the AI. Most poor results come from missing information, not wrong wording. the more precisely you explain what you want, the better the result. See the glossary entry on prompts for the basic definition.

For a longer walkthrough with before-and-after examples, see the guide How to write AI prompts that work.

The four-part structure

1. Context: who you are, who the audience is, what the situation is.
2. Task: what exactly you want done.
3. Format: length, structure, tone, what the output should look like.
4. Constraints and checks: what to avoid, what to flag, how to handle uncertainty.

What the published guides say

• Number of examples The Anthropic Claude prompting guide recommends including 3-5 examples for best results.
• What examples do The Anthropic Claude prompting guide says a few well-crafted examples, known as few-shot or multishot prompting, improve accuracy and consistency.
• Structure The Anthropic Claude prompting guide says XML tags help Claude parse complex prompts unambiguously when a prompt mixes instructions, context, examples, and variable inputs.
• Research basis The arXiv paper by Tom B. Brown and colleagues reports that scaling up language models greatly improves task-agnostic, few-shot performance.

Before and after

Before

Write a summary of this report.

After

I am a regional operations manager. The attached quarterly report is going to our VP, who will have five minutes. Summarize it in five bullets: the headline result, the two biggest variances against plan with figures, one risk, and the decision we need from her. Use plain language, no jargon. If a figure is not in the report, write “not in report.”

The second version gets a usable result the first time because the assistant does not have to guess anything.

Techniques that help

Show an example. “Format it like this: [example]” is more reliable than describing the format.

Assign a role for perspective. The Anthropic Claude prompting guide says setting a role in the system prompt focuses Claude’s behavior and tone, and even a single sentence makes a difference. “Review this proposal as a skeptical CFO” tends to produce sharper critique than “review this.”

Ask it to ask you. “Before you start, ask me up to five questions that would improve the result.” This works well when you are unsure what to include.

Break large tasks up. Outline first, review, then draft each section.

Ask for the reasoning or sources. For factual work: “For each claim, say where in the document it comes from.”

Give permission to say no. “If you cannot tell, say so” cuts down on invented answers. More in how to stop AI hallucinations.

Iterating

Treat the first answer as a draft. Say specifically what is wrong: “Too formal. Cut the second paragraph. Lead with the deadline.” Vague feedback like “make it better” gives vague improvement.

If a conversation gets long and confused, start a new one and paste the best version of your instructions.

Common mistakes

• One-line requests for complex tasks.
• Pasting a huge document with no question.
• Not saying who the audience is.
• Accepting the answer without checking.
• Rewriting the same prompt from scratch each time instead of saving it.

Build a prompt library

When a prompt works, copy it into a document with a note on when to use it. Re-test saved prompts from time to time, because a model update can change the results.. After a month you will have a set of reliable tools. Learn AI Fast teaches a framework called CRAFT (Context, Request, Alternatives, Feedback, Transfer) for getting dependable output from any AI tool.

Related

Two applications of these ideas are emails and writing in your own voice.

Sources

• Anthropic, Claude prompting best practices: “Include 3–5 examples for best results”
• Anthropic, Claude prompting best practices: “A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency”
• Anthropic, Claude prompting best practices: “XML tags help Claude parse complex prompts unambiguously”
• arXiv, Language Models are Few-Shot Learners (Brown et al.): “scaling up language models greatly improves task-agnostic, few-shot performance”
• Anthropic, Claude prompting best practices: “Setting a role in the system prompt focuses Claude’s behavior and tone for your use case. Even a single sentence makes a difference”