Prompting Claude for Marketers: Roles, Context and Examples

#01Why does prompting still matter in 2026?

Models are much better at guessing what you mean than they were two years ago. But Anthropic's own guidance says its current models are built to follow instructions precisely, and it advises being explicit about what you want rather than expecting the model to infer it (Anthropic). Precision is a gift when you have a clear brief and a problem when you don't. A vague prompt gets a competent, generic answer. A specific one gets work you can use.

Anthropic suggests a simple "key test": 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. For marketers, the colleague test catches most problems before you waste a message.

#02How should you set a role?

A role tells Claude whose perspective to take. Anthropic notes that even a single sentence of role setting makes a difference (Anthropic).

Weak: "You are a marketing expert."

Better: "You're a senior B2B content editor who writes for finance managers at mid-sized companies. You value plain language, concrete examples and short paragraphs."

The better version does two jobs. It sets expertise, and it tells Claude about the audience and the standards to apply. You can also assign the reader's role rather than the writer's: "Read this landing page as a sceptical CFO who's seen too many software pitches. What would make you stop reading?" Reader roles are one of the most useful editing tricks in this track.

#03What context does Claude need?

More than you think, and the reason behind it. Anthropic's guide gives a small but telling example: "NEVER use ellipses" works less well than explaining that the response will be read aloud by a text-to-speech engine, so ellipses won't be pronounced. When Claude knows why a rule exists, it can apply the spirit of it to cases you didn't foresee.

For marketing, context usually means:

• The audience: who they are, what they know and what they're sceptical about.
• The goal: what the reader should think, feel or do afterwards.
• The situation: where the copy will appear, what comes before and after it, and what's been tried already.
• The facts: product details, prices, proof points and anything Claude must not get wrong.
• The reason for each constraint: "Keep it under 150 characters because the email preview cuts off there."

A good rule: if you'd spend five minutes explaining something to a freelancer, spend two minutes writing it for Claude. It's still far quicker than editing generic output.

#04How do you use examples well?

Examples are the most reliable way to get a style or format right. Anthropic recommends three to five examples, made relevant to your use case, varied enough that Claude doesn't copy a single pattern, and wrapped in <example> tags so they're clearly separate from instructions (Anthropic).

Two points matter for marketers. First, use your best real work, not invented samples. Your three highest-performing subject lines tell Claude more about your voice than a paragraph of adjectives. Second, vary the examples. If all three examples are questions, you'll get twenty questions back. Include different structures so Claude learns the voice, not the template.

You can also ask Claude to critique your examples: "What do these three have in common? What makes them work?" Then use its answer as part of your brief.

#05How do you state constraints?

Plainly and positively where you can. Anthropic's guidance suggests telling Claude what to do rather than what not to do, and notes that its newer models respond well to calm instructions; you don't need to shout with "CRITICAL" or "MUST" in capitals, which can make it overreact (Anthropic).

Instead of "Don't use jargon", try "Use words a finance manager would use in an email to a colleague." Instead of "Don't be salesy", try "Write as if you're recommending the product to a peer, and include one honest limitation."

Hard constraints should still be clear: character limits, required disclaimers, product names spelt exactly, banned claims. List them together in one place so Claude, and you, can check them.

#06How should you structure a long prompt?

Use tags and put things in the right order. Anthropic recommends XML-style tags such as <instructions>, <context> and <document> to separate the parts of a complex prompt. There's nothing magic about the tag names; they just stop Claude mixing up your instructions with the material it's working on.

Order matters for long inputs. Anthropic's guide says to put long documents near the top and your question at the end, and reports that queries at the end can improve response quality by up to 30% in its tests, especially with complex, multi-document inputs. For marketers, that means pasting the transcript, brand guide or research first, then asking.

The guide also suggests asking Claude to pull out relevant quotes first, in <quotes> tags, before answering. That's very useful for research and fact-checking because you can see what the answer is based on.

#07What does a good marketing prompt template look like?

Here's a template you can save to your prompt library. Replace the brackets.

<context>
I'm the [role] at [company]. We sell [product] to [audience],
who care most about [priority] and are sceptical of [objection].
This copy will appear [where], read by people who [situation].
</context>

<materials>
[Paste brand guide extract, product facts, research]
</materials>

<examples>
<example>[Your best real example 1]</example>
<example>[Your best real example 2]</example>
<example>[A different format that still sounds like us]</example>
</examples>

<instructions>
Write [deliverable] that makes the reader [goal].
Constraints: [length, format, required phrases, claims to avoid]
and why each matters: [reason].
Before writing, list the three facts from <materials> you'll rely on.
Then give me [number] options, each with a one-line rationale.
</instructions>

#08How do you ask for action rather than suggestions?

Anthropic notes a subtle point: if you ask "Can you suggest some changes?", Claude may suggest rather than make them, because its newer models take wording literally. If you want Claude to edit the document, say "Change this…" or "Rewrite this…" (Anthropic). The reverse also works: if you want options without changes, ask for suggestions.

The same applies to format. If you want flowing prose rather than bullet points, say so and explain why, for example "This is a blog section; write it as paragraphs." Anthropic also notes that the style of your prompt influences the style of the answer, so a prompt written in prose tends to get prose back.

#09How do you iterate without wasting messages?

When a result is close but not right, edit your original prompt and resend it rather than adding corrections. That's cheaper on usage, as lesson 2 explained, and it keeps the conversation clean.

When the result is right, ask Claude to explain what worked: "Write the instructions I should give you to get this result first time." Save that to your library. Over a few weeks, you'll build prompts that work reliably without editing.

#10From MentionX data: why your prompts should mirror buyer questionsOriginal data

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

When you use Claude to research your market, phrase test questions the way buyers do. MentionX's aggregates show how much the question type changes whether brands appear:

Our view: a prompt about "the best accounts-payable software" and one asking "Ledgerline vs a competitor" will give you very different pictures of your market. Write research prompts in several intents, and see our guide to building a prompt set for how many to use.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

#11Worked example: rewriting a weak prompt

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline's lifecycle owner, Sam, starts with: "Write a welcome email for new customers. Make it friendly and not too long."

Claude's draft is fine and forgettable. Sam rewrites the prompt using the template. The context explains that new customers are usually finance managers who've just been pushed onto the tool by their CFO and are slightly resentful. The materials include the three setup steps and the support link. The examples are the team's two best-performing onboarding emails. The constraints say 120 words maximum because most open on mobile, and "no exclamation marks, because our readers find them insincere".

The second draft opens with "You've got a lot on. Here's the five-minute version." Sam keeps it, then asks Claude to write the instructions it would need to get that result first time, and saves the answer as a prompt.

#12What should you practise before the next lesson?

Take one piece of copy you need to write this week. Write a weak prompt and a strong one using the template, and compare the results side by side. Note which part of the strong prompt made the most difference; for most people, it's the examples. Start your prompt library with that template and the result.

Next: Projects and instructions for brand voice, where you'll stop pasting the same context into every chat. Or return to the Claude for Marketers track.

#13What changed in prompting recently?

Two things worth knowing. First, Anthropic's current guide notes that prefilling Claude's response, a popular trick for forcing a format, is no longer supported on its newest models; use clear format instructions or structured outputs instead. Second, its newer models are more responsive to instructions than earlier ones, so prompts written for older models with lots of emphatic warnings can now cause over-correction. If an old prompt suddenly produces stiff or overcautious copy, soften the language before rewriting anything else.

#14How do you prompt differently for research and for writing?

Writing prompts need voice, examples and constraints. Research prompts need scope, sources and a way to check the answer. Mixing the two in one prompt is a common reason for weak output, so separate them where you can.

For research, tell Claude what decision the research supports, what time period and market matter, and what kind of sources you trust. Ask it to cite every claim and to say clearly when it couldn't find evidence. Anthropic's guide recommends grounding long-document answers in quotes first, which works well for research too: ask Claude to list the passages it's relying on before it summarises them (Anthropic).

I'm deciding whether to create a comparison page for [category].
Search for pages published in the last 12 months that compare
[category] tools for [audience]. For each, give the URL, publisher,
date and the criteria it uses. Quote the sentence that shows each
criterion. If you can't find a date, say so. Don't summarise
pages you couldn't open.

For writing, start a fresh chat and give Claude only the research it needs, not the whole conversation. That keeps the draft focused and stops early dead ends creeping into the copy.

Our view: the most useful single habit is ending research prompts with "If you're not sure, say so." It makes uncertainty visible, which is exactly what you need before anything goes in front of customers.

#15How do you use XML tags in long prompts?

Anthropic's prompting guide recommends XML tags when a prompt mixes instructions, context, examples and inputs: wrapping each type of content in its own tag, such as <instructions>, <context> or <input>, "reduces misinterpretation" (Anthropic). The tag names aren't special. What matters is that Claude can tell your brand guide apart from the draft it's meant to edit, and the examples apart from the task.

<context>
We sell [product] to [audience]. Positioning: [one line].
</context>

<brand_rules>
[paste rules]
</brand_rules>

<examples>
<example>[a past email you liked]</example>
<example>[another]</example>
</examples>

<task>
Write a launch email for [feature] using the context and rules.
Match the tone of the examples but don't copy their structure.
</task>

<output_format>
Subject line options (3), preview text, body under 150 words.
</output_format>

This matters most when you paste long material. Without tags, a sentence inside a pasted customer review ("make it shorter!") can be mistaken for an instruction to Claude. You can also refer to tags in the task: "Use only facts in <product_facts>."

#16Which prompt templates are worth saving?

Most marketing teams reuse the same handful of prompts. Save these as text snippets, project instructions or a skill, and fill in the brackets each time.

Critique before rewrite. Ask for the diagnosis before the fix.

Here's a draft [landing page / email / post]. Before changing
anything, list the 5 biggest problems for a reader who is
[audience] and [situation]. Rank them by impact. Then rewrite
fixing only the top 3, and show what you changed.

The sceptical buyer. Useful for any persuasive copy.

Read this as a sceptical [job title] who has seen many vendors
make the same claim. List every sentence you wouldn't believe,
and say what evidence would change your mind.

Positioning check.

Here's our positioning statement and three competitor homepages
(labelled Competitor A-C). For each of our main claims, say
whether a competitor makes the same claim, how they phrase it,
and whether ours is meaningfully different. Be blunt.

Brief from a messy thread.

Below is a Slack thread and meeting notes about [project].
Turn them into a one-page brief: objective, audience, key
message, deliverables, deadline, owner, and open questions.
List anything people disagreed about as an open question.
Don't fill gaps with assumptions; mark them [unknown].

Format conversion.

Turn this [webinar transcript / long post] into [format].
Keep the speaker's strongest phrases verbatim and mark them
with quotation marks so I can check them.

#17What should you do when Claude's output is off?

Most problems map to a missing piece of the prompt. Here's a quick guide.

If a chat has gone on a long time and quality drops, start a fresh one with a summary of what you've agreed. Long conversations carry earlier drafts and abandoned directions that can pull the output back. Asking Claude to "summarise the decisions we've made so far as a brief I can paste into a new chat" is the fastest way to restart cleanly.

#18Common questions

What is the best prompt structure for Claude? Context and long materials first, then examples, then clear instructions and constraints at the end, separated with simple XML-style tags.

How many examples should I give Claude? Anthropic recommends three to five varied, relevant examples wrapped in example tags.

Should I use capitals like MUST or NEVER? Usually not. Anthropic says its newer models respond well to calm, clear instructions, and shouting can cause over-correction.