To write a good prompt for Claude, say plainly what you want, tell it why, describe who the result is for and what shape it should take, and paste any long material before your question. Anthropic's own guide sums up the approach in one line: treat Claude as "a brilliant but new employee who lacks context on your norms and workflows" (Anthropic, prompting best practices). The rest of this post turns that guide into habits for everyday Claude.ai use, three before-and-after prompts for ordinary jobs, and a short section for developers using the API or Claude Code.
How does Anthropic say to prompt Claude?
Anthropic keeps one living reference for prompting its current models, and its general advice comes down to seven habits. Most of them apply as much to a chat in Claude.ai as to an app built on the API.
- Be clear and direct about what you want.
- Give the reason behind your instructions.
- Show examples of the output you want.
- Separate instructions from material with tags or headings.
- Let Claude think before it answers on hard problems.
- Put long documents first and the question last.
- Describe the output you want, including the tone.
Each one is below, with what it looks like in a normal chat.
Be clear and direct
Anthropic's "golden rule" is to show your prompt to a colleague with minimal context on the task. If they would be confused, Claude will be too. The guide also says that if you want Claude to go "above and beyond," you have to ask for it explicitly, because it won't infer that from a vague prompt.
In practice, name the deliverable ("a 150-word email," "a table with four columns"), the audience and any limits. When order matters, number the steps. And ask for the action you want: Anthropic notes that "can you suggest some changes" sometimes gets suggestions when you wanted the changes made, so write "edit this paragraph" if that's what you mean.
Tell Claude why
A rule with a reason works better than a bare rule. The guide's example: "NEVER use ellipses" is less effective than explaining that the response will be read aloud by a text-to-speech engine that won't know how to pronounce them. With the reason, Claude can apply it to cases you didn't list.
For everyday prompts, the "why" is usually your situation. "Keep it short" is a rule. "Keep it short, my manager reads these on her phone between meetings" lets Claude decide what to cut.
Show an example
Examples are one of the most reliable ways to steer format, tone and structure, according to Anthropic, and the guide recommends three to five for best results. Make them varied, and wrap each one in <example> tags so Claude can tell an example from an instruction.
Anthropic's blog adds a warning: newer Claude models pay close attention to examples, so an example should show only behavior you want copied (Claude blog, best practices for prompt engineering). If all three sample emails open with "Hope you're well," expect the fourth to.
Separate the parts with tags or headings
When a prompt mixes instructions, pasted material and examples, tags stop Claude from confusing one for another. Anthropic suggests descriptive names such as <instructions>, <context> and <document>, used consistently. You don't need a special syntax: any tag name that describes its content works.
For a short request, this is overkill. The same Anthropic blog post says clear headings and whitespace often do the job for simpler prompts.
Let Claude think first
For problems with several steps, such as a budget, a schedule or a tricky decision, ask Claude to work through it before giving the answer. Anthropic's blog gives a simple version, "Think step-by-step before you write the email," and a structured one that asks for the reasoning inside <thinking> tags. If your Claude.ai plan offers extended thinking, you can switch it on instead of adding the instruction.
Put long documents first and the question last
This is the most Claude-specific habit, and the one with a number attached. For long inputs, Anthropic recommends placing the documents near the top of the prompt, above your question and instructions, and says queries at the end "can improve response quality by up to 30 percent in tests," especially with several documents (Anthropic, long context prompting).
The same section suggests asking Claude to pull out the relevant quotes first and then do the task. That keeps it anchored to what the document says. In Claude.ai, uploading a file and then typing your question below it already gives you the right order.
Describe the output you want
Anthropic's formatting advice is to describe the output you want. "Do not use markdown" works less well than "write in smoothly flowing prose paragraphs." The same goes for tone: "warm and direct, like a note to a colleague you like" gives Claude a target, where "don't be stiff" only rules one thing out.
It also helps to give Claude permission to say it doesn't know. Anthropic's blog suggests a line like "If the data is insufficient to draw conclusions, say so rather than speculating."
Before and after: three everyday Claude prompts
These are prompts typed the way most people type them, each followed by my rewrite using the habits above.
Summarising a long report
Before:
summarise this [pasted report]
After:
<report>
[pasted report]
</report>
I'm presenting this report to our board on Thursday. They care about cost and risk, not methodology. First, quote the three or four passages that matter most for cost and risk. Then write a summary of under 200 words based on those quotes. If the report doesn't answer something the board will ask, tell me instead of guessing.
The report now comes first, the reason (a board that cares about cost and risk) tells Claude what to keep, and the quotes step keeps the summary tied to the text.
Asking your manager for more time
Before:
write a message to my boss asking for more time on the project
After:
Write a Slack message to my manager asking to move the client report deadline from Friday to next Wednesday. The delay is because the client sent their data four days late. I want to come across as on top of it, so one short apology at most. Offer to send the first half by Friday. Three or four sentences.
The rewrite gives Claude the dates, the reason for the delay and the tone, so it doesn't have to invent any of them.
Choosing between two options
Before:
which job offer should i take
After:
Help me choose between two job offers. Offer A: product analyst at a 40-person startup, £42k, 20-minute commute, hybrid. Offer B: analyst at a large bank, £48k, one-hour commute, in the office four days a week. What matters most to me over the next two years is learning fast; salary matters less as long as I can cover £1,200 rent. Compare them in a table on pay, growth, commute and risk, then tell me which you'd pick and why. Ask me first if anything you need is missing.
Claude can now weigh the offers against what you care about, and the last line covers anything you forgot to mention.
Save the context you repeat
If you type the same background into every chat, move it into a pre-prompt. Claude.ai has "Instructions for Claude" in Settings, which apply to all your conversations, and project instructions, which apply to every chat inside one project (Claude Help Center, personalization features). Put your role and standing preferences in the first and topic-specific material in a project. There's more on how these layers work in what a pre-prompt is.
Prompting Claude through the API or Claude Code
Developers get the same advice with a few extra controls.
Set a role in the system parameter. Anthropic's guide says even a single sentence of role in the system prompt focuses Claude's behavior and tone for your use case.
Don't rely on prefill with current models. Anthropic's guide says that starting with Claude 4.6 models, a prefilled last assistant turn is no longer supported and returns a 400 error (Anthropic, migrating away from prefilled responses). To force a format, use structured outputs or tell Claude directly, for example "Respond directly without preamble."
Let effort control thinking. Recent models use adaptive thinking, and the guide recommends moving budget control to the effort setting instead of a fixed thinking budget.
Keep CLAUDE.md short. In Claude Code, the documentation suggests keeping each CLAUDE.md under 200 lines, because longer files consume more context and reduce adherence (Claude Code docs, memory).
The rest of the advice is the same for code as for prose: explain the goal, show an example of the output, and put long inputs before the instruction.
Getting the rewrite done for you
The habits above take a minute per prompt, which is the minute most of us skip. PrePrompt is a Chrome extension that adds a Rewrite button next to Send on Claude, ChatGPT, Gemini and Perplexity. It launches soon with 30 free rewrites a month, no card required; join the waitlist to hear the day it goes live. It leaves a clear prompt alone, and with a long one it organises your own words instead of rewriting them.
TL;DR
- Treat Claude like a capable new colleague with no context: say what you want, who it's for, the format and the limits.
- Give reasons for your rules, and show three to five varied examples when format matters.
- Use tags or headings when a prompt mixes instructions with pasted material.
- Put long documents first and the question last; Anthropic reports up to 30% better responses in tests.
- Developers: set a role in the system prompt, replace prefill on Claude 4.6 and later, and keep CLAUDE.md under 200 lines.
About the author. Tanuj writes PrePrompt's fundamentals and how-to guides and works on the product. PrePrompt is a Chrome extension that rewrites prompts before you send them, and its rewrites run on a Claude Haiku model, so Anthropic's prompting guide is required reading for him.
Sources: Prompting best practices (Anthropic); Best practices for prompt engineering (Claude blog, 2025); Understanding Claude's personalization features (Claude Help Center); Memory (Claude Code docs).
Related reading: what a pre-prompt is, what a vague prompt is, with examples and what a good prompt looks like.
