Claude Prompts for Marketing Teams
Twelve Copyable Prompts for Strategy, Content and Reporting, Built the Way Anthropic Recommends
Short answer
A good Claude prompt for marketing names the job, puts your material in labeled tags and shows the output you want. It also says what to leave out. Anthropic recommends clarity, examples, XML structure and role prompting. The twelve prompts below apply that to positioning, reviews, drafting, ads, landing pages and reports.
1. What makes a Claude prompt work for marketing?
Four parts: a clear job, the material in labeled tags, the exact output shape, and one or two examples. Anthropic documents all four in its prompting best practices, and each one removes a way for the draft to go generic.
Anthropic keeps a living reference on prompting its models that covers clarity, examples, XML structuring, role prompting, thinking and prompt chaining [s1]. For marketing work, those collapse into four parts that every prompt on this page uses.
| Part | What it looks like | What goes wrong without it |
|---|---|---|
| The job | One sentence naming the task and who the output is for | Claude guesses the audience and writes for everyone |
| The material | Your positioning, reviews or data inside tags such as <reviews> | Claude fills the gap from memory of your category |
| The output shape | Format, length, number of options, what to include per item | An essay when you needed a table |
| Examples and exclusions | One good example, a list of words or claims to avoid | The category average voice, and claims you cannot back |
Tags matter more than they look. Wrapping each piece of material in its own named tag lets Claude tell your positioning apart from the reviews you pasted, and lets you refer to it by name: "use only the claims inside <proof>".
2. What is the reusable wrapper for every marketing prompt?
A fixed frame of role, context tags, task, output format and rules. Fill the tags, change the task line, and the rest stays the same. Store it in a Claude Project so the context tags fill themselves.
Copy this frame once. Every prompt below is a task line and output block dropped into it.
- Role. You are the marketing lead at [company], a [category] company for [audience]. You write plainly and never claim anything that is not in the material below.
- Context tags. <positioning>[paste positioning statement]</positioning> <icp>[paste ideal customer profile]</icp> <voice>[paste voice rules and deny list]</voice> <proof>[paste numbers, quotes and case facts you are allowed to use]</proof>Inside a Claude Project these live in project knowledge, so you only paste the material specific to the task.
- Task. One sentence: what you want, for which channel, and what it should change in the reader.
- Output format. Say exactly what comes back: a table with named columns, three options labeled A to C, or a draft under a word count.
- Rules. Use only claims from <proof>. If something needed is missing, list it under "Missing" instead of inventing it. No rhetorical question openers. Follow <voice>.
3. Which Claude prompts work for positioning and strategy?
Prompts that make Claude argue rather than agree: a skeptical-buyer pressure test, an ICP built from real reviews, a message house from your positioning, and an objection library with proof attached.
Four strategy prompts. Each assumes the wrapper above.
- Positioning pressure test. Task: "Argue against the positioning in <positioning> as a skeptical buyer matching <icp>. List the five weakest claims, why each would not convince you, and what evidence would." Output: a table with Claim, Why it fails, Evidence that would fix it.
- ICP from reviews. Material: <reviews>[paste 50 to 200 of your own reviews]</reviews>. Task: "Group reviewers by the job they hired us for. For each group give the count, the trigger that made them look, the words they use for the problem, and two verbatim quotes." Output: one block per group, largest first.
- Message house. Task: "Build a message house from <positioning> and <proof>: one umbrella message, three pillars, and for each pillar two proof points taken only from <proof>." Output: a nested list, plus a Missing list for any pillar without proof.
- Objection library. Material: <objections>[paste objections from sales calls or reviews]</objections>. Task: "For each objection write a two-sentence answer using only <proof>, and label whether the objection is about price, risk, effort or fit." Output: a table with Objection, Type, Answer, Proof used.
The ICP and message house prompts have their own templates in the Playbook guides linked below if you want the blank structures as well.
4. Which Claude prompts work for drafting content?
Prompts that ask for different angles rather than different versions, that tie each draft to one belief, and that hand Claude your own posts as examples. One brief per piece beats a batch request every time.
Four drafting prompts:
- Three angles on one brief. Task: "Write three LinkedIn posts for <icp> that each try to change the belief [belief] using one proof point from <proof>. Each must use a different angle: a customer story, a counterintuitive claim, and a how-to." Output: A, B, C, each under 150 words, with the angle named on top.
- Repurpose one idea. Material: <source>[paste a blog post or webinar notes]</source>. Task: "Turn <source> into an X post, a LinkedIn post and an Instagram caption. Keep the single most specific detail in all three." Output: three drafts labeled by channel.
- Voice rules from examples. Material: <posts>[paste ten posts you would publish again]</posts>. Task: "List what these posts have in common as rules a reviewer could check: sentence length, openers, person, claim style, words never used." Output: six to ten numbered rules.
- Ad variants from the message house. Task: "Write four ad variants for Meta, one per pillar in the message house, each with a headline under 40 characters and primary text under 125 characters." Output: a table with Pillar, Headline, Primary text, Proof used.
5. Which Claude prompts work for analysis and reporting?
Prompts that supply the numbers and ask for the narrative, never the reverse. Claude is good at explaining a table you give it and bad at producing figures it was not given.
Four analysis prompts:
- Review theming with counts. Material: <reviews>[paste reviews]</reviews>. Task: "List complaint themes with the number of reviews mentioning each and two verbatim quotes per theme. Put anything that fits no theme under Other with a count." Output: a table sorted by count.
- Landing page critique. Material: <page>[paste page copy]</page>. Task: "Read <page> as a buyer from <icp> arriving from an ad. Where do you stop reading, what is unclear, and what question is left unanswered?" Output: numbered points in page order.
- Monthly report narrative. Material: <numbers>[paste this month and last month by channel]</numbers>. Task: "Write a five-sentence summary for a founder: what moved, the likely reason if the data shows one, and one decision to make. Do not state any number that is not in <numbers>." Output: prose, then a Missing list.
- Interview me first. Task: "Before writing the Q4 plan, ask me up to eight questions you need answered. Ask them one at a time and wait for each answer." This is the prompt to use when you are not sure what material the task needs.
6. How should a team chain and store these prompts?
Chain them in order, one output feeding the next, and store the stable parts in a Claude Project or a skill. The wrapper and your material stay fixed; only the task line changes.
Prompt chaining is in Anthropic guidance for a reason: one prompt doing five jobs does all of them badly [s1]. A useful chain for a small team is reviews to ICP, ICP to message house, message house to ad variants, and each step reviewed by a person before the next one runs.
Keep standing context in a Project, which holds its own knowledge files and instructions [s2]. Once a prompt is stable, a skill packages it as a set of instructions Claude loads when relevant, and simple skills are written in plain Markdown with no code [s3].
If your competitor data is tracked in Oppira, the same prompts work over connected data rather than pasted text: through the MCP server on Basic and Pro, Claude can read the reviews, posts and playbook the prompts refer to.
Key Takeaways
Four parts in every prompt
Job, material in tags, output shape, examples and exclusions. Each one removes a way to go generic.
Tags separate your material
Named tags let Claude tell positioning from reviews and let you restrict claims to what is inside <proof>.
Ask for a Missing list
Making Claude list what it needed and did not have turns invention into a visible gap.
Ask for angles, not versions
Three different angles on one brief are more useful than three rewrites of the same idea.
Supply numbers, ask for narrative
Claude explains a table well and invents figures badly. Never ask it for numbers you did not give it.
Chain, then store
Run prompts in sequence with a review between steps, and keep the stable parts in a Project or skill.
Frequently Asked Questions
Sources
- Prompting best practices Anthropic, September 2026.Covers clarity, examples, XML structuring, role prompting, thinking and prompt chaining.
- What are projects? Claude Help Center, September 2026.
- What are skills? Claude Help Center, September 2026.
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. See your market clearly by tomorrow.