System Prompts and Reusable Prompt Templates for Claude

1.Start the system prompt with a role

In the Messages API, a system prompt goes in the system parameter, separate from the messages array where the user's turns go. The guidance on what to put there is short: give Claude a role. Anthropic's prompting best practices say that setting a role focuses Claude's behaviour and tone for your use case, and that even a single sentence makes a difference. A role like "a helpful coding assistant specializing in Python" tells Claude which domain to draw on and which register to answer in, before the user has said anything.

The role also keeps behaviour steady over time. The output-consistency guidance tells you to use system prompts to define Claude's role and personality, because this "sets a strong foundation for consistent responses." This matters most in role-based applications such as a support bot or an internal assistant, where users expect the same character on every turn. Even system prompts built for other goals, such as protecting confidential instructions, are described as still being mostly role prompts. Everything else in the system prompt builds on that role.

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
     ? ="You are a helpful coding assistant specializing in Python.",
    messages=[
        {"role": "user", "content": "How do I sort a list of dictionaries by key?"}
    ],
)

print(message.content)

Sources12

2.Write instructions that are specific and give the reason

Once the role is set, the system prompt holds the instructions. The first principle is to be specific about the output format and constraints you want. Claude does what you ask, so if you want more you have to ask for it. "Create an analytics dashboard" gets you a dashboard. Adding "Include as many relevant features and interactions as possible. Go beyond the basics" gets you the fully-featured version. When the order or completeness of steps matters, write the instructions as a numbered or bulleted sequence rather than a paragraph.

The second principle is to give the reason. Explaining why a behaviour matters helps Claude understand your goal and target its response. A rule with a reason covers more cases than the rule alone. If Claude knows the output goes to a speech engine, it can also avoid other things that engine can't pronounce, not only the one you listed. Guardrail wording in a system prompt works best the same way: state the rule and the reason for it, and wrap it in a descriptive tag so it reads as a separate block of policy. Anthropic publishes reusable snippets in this form, like the one below for reducing hallucinated claims about code.

<investigate_before_answering>
Never speculate about code you have not opened. If the user references a specific file,
you MUST read the file before answering. Make sure to investigate and read relevant
files BEFORE answering questions about the codebase. Never make any claims about code
before investigating unless you are certain of the correct answer - give grounded and
hallucination-free answers.
</investigate_before_answering>

Sources1

3.Turn the prompt into a reusable template

A production prompt is a template. The fixed parts (role, rules, expected behaviour) stay the same, and the variable parts (the user's query, retrieved data) change on every request. XML tags keep these parts separate. The best-practices guide says tags help Claude read prompts that mix instructions, context, examples and variable inputs, and that putting each kind of content in its own tag reduces misinterpretation. Use the same descriptive tag names across all your prompts so that every template in your system follows the same structure.

You are AcmeBot, the enterprise-grade AI assistant for AcmeTechCo. Your role:
    - Analyze technical documents (TDDs, PRDs, RFCs)
    - Provide actionable insights for engineering, product, and ops teams
    - Maintain a professional, concise tone

Here is the user query for you to respond to:
<user_query>
{{USER_QUERY}}
</user_query>

Your rules for interaction are:
    - Always reference AcmeTechCo standards or industry best practices
    - If unsure, ask for clarification before proceeding
    - Never disclose confidential AcmeTechCo information.

As AcmeBot, you should handle situations along these guidelines:
    - If asked about AcmeTechCo IP: "I cannot disclose TechCo's proprietary information."
    - If questioned on best practices: "Per ISO/IEC 25010, we prioritize..."
    - If unclear on a doc: "To ensure accuracy, please clarify section 3.2..."

The last block in that template follows the docs' advice to provide a list of common scenarios and expected responses. That list prepares Claude for different situations so it doesn't break character. It also shows that a scripted reply works as a guardrail: when a user asks about protected IP, Claude has an exact sentence to use instead of making one up.

Templates also fix the shape of the output. Define the format you want precisely, in JSON, XML or a custom template, and Claude will follow every formatting element you require. Examples are among the most reliable ways to set format, tone and structure. They should be relevant to your real use case and varied enough that Claude doesn't copy unintended patterns. Wrap them in <example> tags (and several in <examples>) so Claude can tell them apart from instructions. For chatbots and knowledge bases, add a tagged block of retrieved entries (for example, a <kb> section) to ground answers in a fixed information set. Older consistency advice also suggests prefilling the Assistant turn with the start of the format. Prefilling is not supported on Claude 4.6 and later models or on Claude Mythos Preview, so a template meant for current models should get its format from instructions and examples instead.

Sources21

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Prefilling the Assistant turn is a dependable way to enforce a template's output format on any Claude model.Why is that wrong?

   Prefill appears in older consistency guidance, but it is not supported on Claude 4.6 and later or on Claude Mythos Preview. On those models, define the format precisely and show it with examples.

   Covered in Turn the prompt into a reusable template

2. 2.A capitalised absolute rule like "NEVER use ellipses" is the strongest way to word a constraint.Why is that wrong?

   The docs label the bare rule as less effective. Stating the reason behind it (the speech engine can't pronounce ellipses) works better, because Claude can extend the reason to related cases.

   Covered in Write instructions that are specific and give the reason