How to write an effective prompt for Claude
The quality of a Claude answer depends mostly on how clear the request is. An effective prompt is not a magic formula: it is an instruction that gives context, states the expected result and leaves little room for interpretation. Here is a simple method, in a few habits, to move from vague requests to precise answers, whether you write on claude.ai or through the API.
In short: Give the context (who, for whom, for what purpose), specify the expected format and tone, show an example if the result is hard to describe, and iterate rather than start over. In short: describe the real situation, say exactly what you want as output, and refine through successive feedback in the same conversation.
Give context and a role
The first mistake is asking for too little. Instead of write a follow-up email, specify the situation: who you are writing for, to whom, for what purpose, with what history. You can also assign Claude a role (you are a sales lead following up with a client after two weeks of silence), which anchors the tone and vocabulary. Context is the difference between a generic answer and one you can use as is. The higher the stakes, the more you gain by describing the real situation in two or three sentences before the request.
Specify format and constraints
State explicitly what you expect as output: a paragraph or a list, a table, a word count, a register, a response language. Clear constraints avoid back-and-forth: reply in five points maximum, professional but warm tone, no superlatives, in English. You can also state what not to do (do not invent any figures, do not propose a paid solution). For outputs reusable by another tool, ask for a precise structured format (a table with these columns, or JSON with these fields).
Show an example (few-shot)
When a result is hard to describe, show it. Pasting an example of what you expect, or a before/after, is often better than a long description: Claude uses it as a style and structure template. This technique, called few-shot, is especially useful to reproduce a brand tone, an ad format or a recurring layout. One or two well-chosen examples are usually enough. If you have fixed rules (an editorial style guide, terms to avoid), state them once clearly rather than repeating them in every message.
Iterate rather than start over
The best prompt is rarely perfect on the first try, and that is fine. If the answer is not quite right, do not start from scratch: say what is wrong (too long, too formal, a point is missing) and ask for a revision in the same thread. This iterative back-and-forth is often faster than a perfect prompt written in advance. Keep one conversation per topic so the context stays clean. For concrete, copy-ready examples, see our Claude prompts page, and the guide on how to use Claude.
Frequently asked questions
How do I write a good prompt for Claude?
Give the context (who, for whom, for what purpose), specify the expected format and tone, show an example if the result is hard to describe, and iterate rather than start over. In short: describe the real situation, say exactly what you want as output, and refine through successive feedback in the same conversation.
Do prompts need a magic formula?
No. There is no secret formula. What works is simple: clear context, a precise format, an example when useful, and iteration. A well-described instruction almost always beats a wording trick.
What is a few-shot prompt?
It is a prompt where you show one or two examples of the expected result (a before/after, a style template) so Claude reproduces them. Very useful for a brand tone or a recurring format, when describing the result in words would be too long.
Why is my answer not precise?
Most often, the prompt lacks context or format constraints. Add who it is for and the purpose, specify length and tone, give an example, and say what not to do. Then refine through iterations instead of rewriting the whole prompt.
See also: the complete guide to Claude · Claude news in real time
Claude News is an independent publication, not affiliated with Anthropic.