7 golden rules for prompting Claude - the official guide from Anthropic
🚀 Intro
Why do I need these rules?
Working with Claude day to day, I quickly noticed one thing: the quality of my prompts translates directly into the quality of the responses. Sounds obvious? Maybe. But most people learn prompting from random advice on the internet and then get surprised that the AI generates “AI slop” for them - generic, characterless content.
Meanwhile Anthropic - the company that built Claude - published an official guide. These aren’t random tips from Twitter, they’re guidelines straight from the model’s creators. And trust me, they make a difference.
Imagine asking AI to build a dashboard. Without specific guidance you’ll get:
- Purple gradients
- Rounded corners
- A standard, “safe” design
Why? Because the model has a natural “gravity” pulling it toward the most common patterns from its training data. These seven rules will help you escape that gravity.
🎬 Source material
This post is based on this YouTube video, which in turn draws on Anthropic’s official guide to prompting Claude.
✨ Rule 1: Be clear and specific
What’s the idea?
Modern AI models are great at following instructions. The problem is that when you don’t give specific guidance, the model defaults to the “safest” option - what it has seen most often in training data. That’s the “gravity” - it pulls you toward mediocrity.
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Build an analytics dashboard. | Build an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics and create a fully functional implementation. |
| Build a presentation. | Build a professional presentation about our quarterly results. Include thoughtful design elements, visual hierarchy, and engaging animations. |
Key lesson: specificity is your weapon against mediocrity. The more precisely you describe what you expect, the better the results.
🧠 Rule 2: Explain your “why”
What’s the idea?
Claude can reason from context. When you explain the intent behind a request, the model can figure out many things you didn’t say explicitly on its own.
It’s like delegating to an experienced employee - if they understand the goal, they can make better decisions without asking about every detail.
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Write this in a formal tone. | Write this in a formal tone, because it’s going to our board of directors and we need to look credible and professional. |
| Keep it short. | Keep it short, because I’m sending it as a text message and longer messages don’t get read. |
Key lesson: a single sentence explaining “why” can radically change the quality of the response.
📝 Rule 3: Give good examples
What’s the idea?
Modern models imitate examples almost literally. That’s a powerful tool, but also a trap. If your example contains things you don’t want - the model will reproduce them.
The meta-lesson
It isn’t only explicit examples that shape the response - the style of your prompt itself is also a model. If you write:
- Casually and playfully → the response will be similar
- Formally and structured → the response will be similar
Key lesson: treat your prompt as a pattern to be imitated. Write it in the style you expect from the response.
✅ Rule 4: Say what to do, not what to avoid
What’s the idea?
Lots of system prompts are full of negations: “don’t do this,” “avoid that,” “never use…”. The problem is that positive instructions are more effective than negative ones.
Since models are great at following commands, all you need to do is tell them what to do - you don’t have to list everything they shouldn’t do.
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Don’t use markdown in your reply. | Your reply should consist of flowing prose and paragraphs. |
| Make it look nice. | Use clear headings. Bold the key takeaways. Add a summary at the top. |
Key lesson: swap “don’t do X” for “do Y.” Simple.
🎬 Rule 5: Be direct about actions
What’s the idea?
AI models often default to the “safe” option - suggesting instead of acting. If you use cautious language (“you might want to consider…”, “what do you think about…”), the model will be cautious too and won’t change anything.
You want the AI to take action? Say it plainly.
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Could you suggest some changes? | Change this function to improve its performance. |
| What do you think of this proposal? | Edit this proposal - make the benefits clearer and add a call to action at the end. |
Key lesson: use action verbs: “change,” “write,” “edit,” “build.” Avoid: “consider,” “suggest,” “think about.”
🔍 Rule 6: Tap the research potential
What’s the idea?
Claude is great at running research. But to make full use of that potential, you have to give it structure.
A meta-prompt for research
Anthropic published a universal prompt for research tasks:
“Run the research in a structured way. As you gather data, develop competing hypotheses. Track your confidence levels and adjust them as you learn more. Regularly self-critique both your confidence levels and your hypotheses. Break complex questions into smaller, more manageable ones.”
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Research my competitors. | Research my three main competitors in the home-services industry. For each, find their pricing, main services, and customer reviews. Compare them with my company and tell me where I have an advantage. |
Key lesson: give the research structure. Define what, how, and why.
📄 Rule 7: Create professional documents
What’s the idea?
Claude uses special “skills” for creating documents:
- Presentations with animations
- Landing pages
- PDF reports
- Excel spreadsheets
The results are significantly better than they used to be - provided you give the right guidance.
What does it look like in practice?
| Weak prompt | Strong prompt |
|---|---|
| Make me a presentation. | Build a professional presentation on [TOPIC]. Include thoughtful design elements, visual hierarchy, and engaging animations. |
| Write me a report. | Build a monthly report for the team. Summary at the top, sections for each department, charts of progress, tasks for next month. Clean formatting, easy to scan. |
Key lesson: specify the structure, the formatting, the visual elements.
🏁 Summary
Seven rules in a nutshell:
| # | Rule | Key idea |
|---|---|---|
| 1 | Be clear | Specificity is how you escape generic responses |
| 2 | Explain why | Context lets the model reason more |
| 3 | Give examples | The model imitates examples almost literally |
| 4 | Say what to do | Positive instructions > negations |
| 5 | Be direct | Action verbs, not suggestions |
| 6 | Structure research | Give the research frames and goals |
| 7 | Describe documents | Structure, formatting, visual elements |
📌 Closing thought
These rules come straight from Anthropic - the company that built Claude. These aren’t random tips from the internet, they’re official guidelines from the model’s creators.
Remember: the better your prompt, the better the response. It’s worth taking a moment to think the request through before you send it.
Tip: although these rules come from Anthropic and are about Claude, they’re largely universal. Clarity, context, good examples, positive instructions - these work with any large language model: GPT, Gemini, Llama or others.
That’s all! Enjoy it. 🚀
Sources: YouTube video · Anthropic’s official guide