Claude prompting guide: clear tasks and reusable examples

A useful Claude prompt states the task, supplies the information needed to perform it and defines the output. Examples and document structure can make that brief clearer, but they do not establish that the answer is correct.

Anthropic’s prompting reference recommends direct instructions, relevant context, examples and clear separation between instructions and source material. This guide turns those principles into reusable prompts for everyday text tasks.

The templates below are original examples. They are not published test results or promises that a particular wording will produce the same answer every time.

On this page

Start with a task someone could follow

“Improve this” leaves the model to decide what improvement means. For an email, it might shorten the message, make it more formal or change the substance.

Name the job and the constraints instead:

Rewrite the email below for a colleague.
Keep the request, deadline and factual details unchanged.
Use straightforward English and a friendly professional tone.
Return only the revised email.

Email:
[paste your draft]

The task is rewriting; the context is a message to a colleague; the constraints protect the meaning; the output is the email itself.

Read the result against the original. If a date, commitment or qualification changes, reject that change rather than accepting it because the prose sounds smoother.

Give the reason for a constraint

Anthropic recommends explaining context or motivation when it helps the model understand the goal. “Keep it short” can mean anything from one sentence to one page.

An audience and use case make the constraint more concrete:

Summarise these notes for a teammate who missed the meeting.
They need decisions and follow-up actions, not a transcript.
Use the headings Decisions, Actions and Open questions.
For each action, include the named owner and deadline only when
the notes state them. Do not infer missing assignments.

Notes:
[paste meeting notes]

This does not ask the model to guess what happened. It specifies how to organise the evidence provided.

If the notes contain a proposed deadline rather than an agreed one, the summary should preserve that distinction. A prompt cannot make an incomplete source complete.

Use examples to show an output pattern

Examples are useful when the desired format is hard to describe briefly. Anthropic advises making them relevant and varied, including edge cases.

For a small classification task, show what to do when the information is missing:

Extract the delivery date from each message.
Return a table with Message and Delivery date.
Use "Not stated" when the message has no delivery date.
Do not turn a dispatch date into a delivery date.

Examples:
Message: "Your order arrives on 12 October."
Delivery date: 12 October

Message: "We will dispatch your order on 12 October."
Delivery date: Not stated

Messages to process:
[paste messages]

The second example prevents a tempting substitution. It defines an important boundary rather than merely repeating the easiest case.

Review whether your examples cover the distinction the task depends on. Too many nearly identical examples can leave that boundary unclear.

Separate the source from the instructions

Anthropic documents XML tags as a way to separate instructions, context, examples and variable input in a complex prompt. You can use descriptive names without pretending the tags verify the document.

<source>
[paste the policy text]
</source>

<task>
Explain the policy's cancellation terms.
Use only the supplied source.
Separate stated rules from questions the source does not answer.
For each rule, include a short supporting quotation.
</task>

The tags mark where the source begins and ends. The instruction limits the evidence. Neither makes a copied policy current or establishes that the model interpreted it correctly.

Check quotations in the original text. A plausible quotation that is absent from the source is not evidence.

Ask for evidence before a document conclusion

For longer document tasks, Anthropic recommends grounding responses in relevant quotations before carrying out the analysis.

That suggests a useful sequence for comparing two documents:

Compare the two supplied policies on cancellation.
First identify the relevant passages in each policy.
Then make a table with:
- Topic
- Policy A's stated rule
- Policy B's stated rule
- Supporting passage

Do not add a rule that neither policy states.
Do not treat different wording as a substantive difference
unless the cited passages support that conclusion.

A table makes the comparison easier to inspect, but the reader still needs to verify the cited passages. If one source omits a topic, describe the omission without inventing its rule.

Put an actual role to work

Anthropic says setting a role in a system prompt can focus Claude’s behaviour and tone. A role is useful when it names a function, such as an editor protecting factual details.

For everyday use, a brief can still describe that function:

Act as an editor reviewing clarity and factual consistency.
Identify unclear references, unsupported conclusions and
changes of terminology. Suggest the smallest useful revision.
Do not rewrite passages that are already clear.

A role does not supply professional credentials, access to private records or evidence the model has not been given. Asking it to “be an expert” is not a substitute for providing the relevant material.

For the distinction between a reusable instruction and the current task, read what a system prompt is.

Improve the brief from a specific failure

When an answer misses the task, identify the mismatch before adding more instructions. Was the output too long? Did it invent missing details? Did it fail to distinguish a proposal from a decision?

Change the part of the brief that addresses that failure. Keep a small set of representative inputs so that a wording change can be checked against more than the one example that triggered it.

The product, model and supplied context can affect behaviour. These templates reflect documentation consulted October 7, 2026; they should be evaluated for the task in which you use them.

Sources

Anthropic, Prompting best practices, consulted October 7, 2026. The task templates are original worked examples based on the documented techniques.