Claude prompt engineering, in Anthropic's own guide, comes down to a handful of habits: be clear and direct, give the reason behind an instruction, show three to five examples, separate the parts of your prompt, give Claude a role and put long documents above your question. As of Nov 2026 that guide, called Prompting best practices, is Anthropic's living reference for its current models.
Anthropic wrote it for developers, so it talks about system prompts and API calls. You don't need any of that. Every tip below is translated into what you'd type into the Claude chat box or a Project's instructions. If you want general prompt habits that work across any AI tool, my prompt writing guide for small businesses covers those; this page sticks to what Anthropic says about Claude.
Start with the colleague test
Anthropic's best single tip is a test, not a technique: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too."
The guide asks you to think of Claude as a brilliant but new employee who lacks context on your norms and workflows. That's the mindset shift. Most weak prompts aren't badly worded. They're missing what a new starter would need: who it's for, what good looks like, what to avoid and what shape the answer should take.
Anthropic's overview page adds two points before you start. Know what success looks like for your task, and have a first draft prompt to improve, because prompt engineering is the polishing step, not the starting one. It also says not every problem is a prompt problem: sometimes choosing a different model fixes speed or cost more easily.
Ask for exactly what you want
The docs say Claude responds well to clear, explicit instructions, and that if you want more than the task strictly requires, you should ask for it rather than hope Claude works it out. Their example: "Create an analytics dashboard" is less effective than the same request followed by "Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation."
Two more points from that section are easy wins for business writing. Be specific about the format and constraints you want. And when the order of steps matters, give them as a numbered list. A related tip from the formatting advice: tell Claude what to do instead of what not to do. Asking for smoothly flowing prose paragraphs works better than a list of things to avoid.
There's a quieter catch further down the guide. If you ask Claude whether it can suggest some changes, it will sometimes only suggest them. If you want the changes made, say so.
Tell Claude why
This is the tip I'd bet most people skip. Anthropic's example: instead of writing NEVER use ellipses, explain that the response will be read aloud by a text-to-speech engine, so it should never use ellipses because the engine won't know how to pronounce them. The docs then add: "Claude can generalize from the explanation."
The point is that a reason covers cases your rule didn't. Tell Claude your newsletter goes to busy shop owners who read on their phones, and it has something to work from when it decides paragraph length or how much jargon to cut, without you listing every rule.
Show three to five examples
The docs call examples one of the most reliable ways to steer Claude's format, tone and structure, and recommend 3 to 5 for best results. Make them relevant (close to your real use), diverse (so Claude doesn't copy one quirk) and structured, which means clearly marked off from your instructions. You can also ask Claude to check your examples for relevance and variety, or to write more based on the ones you gave it.
For a business, that's three past emails you were happy with, three product descriptions in your house style, or three social posts that did well. Paste them in, label them as examples, then ask for the new one.
Label the parts of a long prompt
Anthropic recommends XML tags to separate instructions, context, examples and input. In plain words, that's a named label in angle brackets at the start and end of each part, such as an instructions tag around your instructions and an example tag around each example. Use the same tag names every time, and nest them when one thing sits inside another, like several documents inside a documents tag.
You don't need this for a three-line request. It earns its keep when your prompt mixes a brief, background, examples and pasted material, and Claude starts treating your example as an instruction.
One sentence of role
The docs say giving Claude a role focuses its behaviour and tone, and that even a single sentence makes a difference. Developers put it in a system prompt. You can put it at the top of your message, or better, in a Project's instructions so it applies every time; my guide to setting up Claude Projects shows where that goes. "You're an editor for a small accountancy firm's client newsletter" tells Claude far more than "write a newsletter".
Long documents first, your question last
For very long inputs, such as a full annual report or several contracts at once, Anthropic says to put the long material near the top, above your question, instructions and examples. It says questions at the end can noticeably improve the answer in its tests, especially when you give it several documents.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Two more tips from the same section: label each document with its source, and ask Claude to quote the relevant parts first before doing the task. The quotes keep it focused on what matters and give you a quick way to check it read the right bits.
A before and after of my own
This is my own example, written to show the techniques side by side, not a result I'm promising. It's a supplier email, the sort of thing most business owners write every month:
Example: before and after
Before: Write an email to our supplier about the late delivery.
After: You're writing for the owner of a small bakery. Write a firm but friendly email to our flour supplier about a delivery that arrived two days late. Keep it to two short paragraphs, because they read email on their phone. Ask for a new delivery date and a named contact. Don't threaten to cancel; we want to keep them.
The after version adds a role, the reason behind the length, the exact ask and one thing not to do, with the reason for that too. It passes the colleague test; the before version doesn't.
Six steps to tune any prompt
- Write the prompt the way you normally would.
- Run the colleague test: would a new starter know what to do with it? Add whatever they'd ask you.
- Add the reason behind your most important instruction.
- Paste in three examples of output you like, clearly labelled.
- If you're including a long document, put it first and your question last.
- Ask Claude to quote the relevant parts before it answers, then check those quotes.
What I'd fix first
I'd fix the colleague test and add one reason before trying anything fancy. Those two are the cheapest changes in Anthropic's whole guide. I'd skip the tag syntax until your prompts get long enough that Claude starts mixing up your instructions and your examples. Once a prompt works, save it in a Project so you're not retyping it, and keep the best ones somewhere your team can find them; my collection of Claude prompts is a good place to borrow starting points.
Anthropic also keeps a separate prompting page for each current model, listing what's different, such as effort levels and response length, and says to read the one for your model first. Model names and features change, so treat this as accurate as of Nov 2026 and check Anthropic's own prompting docs for the latest.
If your team wants prompts that work every time instead of one person's lucky wording, that's the kind of AI enablement I help with. I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
What is the colleague test in Claude prompt engineering?
Anthropic's guide says to show your prompt to a colleague with minimal context and ask them to follow it. If they'd be confused, Claude will be too.
How many examples should I give Claude?
Anthropic recommends 3 to 5 examples for best results, kept relevant, varied and clearly marked off from your instructions.
Where should a long document go in a Claude prompt?
Near the top, above your question and instructions, with the question at the end. Anthropic also suggests asking Claude to quote the relevant parts first.
Do I need XML tags to prompt Claude well?
Not for short requests. Anthropic recommends tags for prompts that mix instructions, context, examples and pasted material, so Claude can tell the parts apart.
Does Claude need a role?
It helps. Anthropic says even a single sentence of role focuses Claude's behaviour and tone. In the chat, put it at the top of your message or in a Project's instructions.
