Almost every marketing team already uses AI. McKinsey's State of AI 2025 survey found that 88% of organizations use AI regularly in at least one business function, yet nearly two-thirds have not begun scaling it across the enterprise.

At the start of 2026, our marketing team at BotsCrew matched that statistic exactly. Everyone used AI, each person alone: personal prompts, private chats, tricks that never left one laptop. In Q2, the whole company moved to one tool, Claude, and we started setting it up as a shared teammate instead of a personal assistant.

The time savings were real, and the numbers are below. But as head of marketing, I'd name a different result first: marketing stopped waiting on other teams. Case studies stopped waiting for engineers to find time. New pages stopped waiting in the developer queue. Research stopped waiting on an afternoon of spreadsheet exports.

Below are five marketing use cases: how each worked before, how it works now, and a prompt you can copy.

What does "Claude as a teammate" mean?

"Claude as a teammate" means Claude works inside a team's real systems under shared rules, instead of in one person's private chat. Picture a very smart new hire who has read every book but knows nothing about your company. Before they can do good work, they need three answers.

  • Instructions: who am I? Claude's role and the rules it follows, such as "take numbers only from approved sources" or "never invent quotes." They rarely change.
  • Context: where am I? Connectors to the systems where work lives. Our marketing team connects Claude to Fathom (call recordings), Jira and Confluence, Google Drive, Gmail, Slack and HubSpot. Connectors run on the Model Context Protocol, and Anthropic's guide to custom connectors shows how to add them.
  • Playbooks: how do things work here? Written steps for recurring tasks. Our rule: anything you do twice becomes a playbook, and the stable ones become skills.

Claude as a teammate

The hardest part, the instructions, Claude can draft itself. We open a new chat in Claude Cowork and send this:

I want to create a Claude Cowork project for [team or task]. Our team works according to the approach on this page [link to wiki page]. Gather information from [call recordings, Slack, and all other available sources] and write a description and instructions for the project using the template from [link to template].

A person reviews the draft, then pastes it into a new shared Project. How we rolled this out across five departments covers the same setup for sales, finance and PeopleOps.

Not every job needs the same container. This is how we choose:

Setup

Best for

Our marketing example

Claude Project

Shared context and rules for one team or task

One project per recurring content type

Skill

A repeatable output in a fixed format

AI-search article research and review

Scheduled task in Claude Cowork

Work that runs on a timer

Morning to-do digest

Claude-built tool

A fix inside a tool you already use

Link-building Chrome extension

Custom AI product

Customer-facing work, live systems, automated decisions

A build for an engineering team, not a project

Use case 1: Case studies from Jira, Confluence and Fathom

How it worked before. For a B2B buyer choosing an AI vendor, a case study with real numbers is the strongest proof we can deliver. Our sales team even builds proposal decks from a bank of approved ones. Yet case studies were our slowest format. The facts lived with the project team, spread across Jira tickets, Confluence pages and client calls, and the people who knew them were busy with client work. Every case study meant interrupting engineers and PMs, then waiting. Gathering, writing and design took 1–2 weeks.

How it works now. With connectors in place, Claude gathers the facts itself. When we needed an urgent case study the flow looked like this:

  1. Claude pulled the project details from Jira and Confluence.
  2. It wrote the draft, which we proofread and edited by hand.
  3. Through Fathom, it found client quotes from calls and sorted them into safe to publish and confidential, with a reason for each.
  4. In Claude Design, we laid it out on the BotsCrew brand system and exported a PDF.

The quote sorting matters more than it looks. A quote that reveals how a client runs internally can't go public. Claude flags it; a person still makes the final call.

Prompt you can adapt:

I'm writing a case study about [project name] for our website. Using Jira, Confluence and our Fathom call recordings, gather: the client's problem, what we built, the tech stack, timeline and measurable results. Then find client quotes from calls that could work as a public testimonial. For each quote, mark whether it is safe to publish or contains confidential details, and explain why. Write the case study as challenge, approach, solution, results. Use only numbers you found in the sources and name the source for each. If a number is missing, write [missing] instead of guessing.

Before → after: 1–2 weeks → about 2 hours, including our proofreading. Nobody on the project team gets pulled into a meeting.

Case studies from Jira, Confluence and Fathom

Use case 2: Website pages with Claude Design and Claude Code

How it worked before. Apart from client referrals, about 90% of our new business comes in through the website. When BotsCrew repositioned around AI consulting, we needed new service and landing pages fast. The blog ran on WordPress, so marketing could publish there, but each post took 2–3 hours of fighting the platform. Every other page was custom-coded and needed a designer, then a developer. Those developers were on client projects, so our pages waited, and SEO work waited behind them.

How it works now. We rebuilt the website with Claude and split the jobs. Sanity, a headless CMS, holds the content, and marketing publishes there directly. Claude builds the front end:

  • Migration. Custom skills describe how old content maps to the new structure, including page styles and custom blog elements. Claude Code moved the pages and the full blog archive; the team reviewed the result.
  • New pages. Copy starts in a Google Doc. We drop it into Claude Design with the BotsCrew design system, then hand the design to Claude Code, which turns it into a production-ready page.
  • Small changes. Instead of a ticket queue, the team writes a short spec. The clearer the spec, the faster the turnaround.

The two prompts we use. In Claude Design, with the website code, the Figma file and the spec attached:

Design a new page for the BotsCrew marketing website based on the .docx spec prepared by the Marketing Department. The page should look modern and match the overall feel and styles of other pages to look consistent.

Then in Claude Code, in Plan mode:

Your task is to implement a new page in the BotsCrew marketing website based on the generated prototype HTML page by converting the prototype into a proper, production-ready page.

Before → after: a custom sales page went from weeks to 2–3 hours. A blog post went from 2–3 hours to about 30 minutes. Designers and developers now review the output instead of producing it, and someone technical still checks the code and deployment before a page goes live.

Website pages with Claude Design and Claude Code

Use case 3: AI-search articles with Ahrefs MCP and skills

How it worked before. B2B buyers increasingly skip the search results page. They ask ChatGPT, Gemini or Perplexity who to hire and get a short list of vendors. If we're not on that list, the buyer may never visit our site, however well we rank. AI assistants favor a clear answer they can lift, sections that make sense out of context, sourced numbers and comparison tables. Our articles started from keywords, and validating which questions to target meant an afternoon of exports and spreadsheets.

How it works now. One writer runs a four-stage process, with Claude connected to Ahrefs through MCP so it pulls search data directly. The writer makes the decisions and owns the text. The stable steps live in a skill, which is how we think about Claude marketing skills: a playbook that runs the same way every time.

  • Research. Claude proposes 10 questions a buyer would type into an AI assistant and checks each in Ahrefs. The writer keeps 5–8. Questions that match prompts we track in HubSpot AEO always stay, because we can measure them.
  • Writing and review. Claude drafts against a fixed structure: the answer on the first screen, self-contained sections, comparison tables, sourced statistics and an FAQ. The writer rewrites it in our voice, adds project data and engineers' opinions, and checks every statistic against its source. Claude then reviews the text against the same structure.
  • Promotion. Claude suggests which published articles to reuse and drafts LinkedIn, YouTube and newsletter versions. On Reddit, only real practitioners from our team answer, under their own names.
  • Measurement. Every month we review AI visibility in HubSpot AEO, plus traffic and leads in Google Analytics.

Prompts from our internal guide, used in order:

I'm writing an article on [topic], aimed at [audience], at the [TOFU / MOFU / BOFU] stage. Give me 10 questions a buyer would actually type into ChatGPT, Gemini, or Perplexity on this topic. Mix definitional, comparative, cost, and "is this right for my situation" phrasings. Write them the way a person prompts, not the way an SEO writes a keyword.

Run these 10 questions through Ahrefs. For each, give me the search volume, keyword difficulty, and the closest-matching terms and search suggestions. Then tell me which 5–8 to keep: the ones with real demand that we can answer better than a generic source. Flag any that are too competitive.

Here are the questions I'm keeping. Draft an outline where each is a self-contained section. Flag which need a comparison table, and which need our own project data to be worth citing.

Before → after: an afternoon of exports → a validated question list in minutes. One writer now covers research, drafting, review and repurposing. Our Claude for marketing guide has the full prompt set.

Use case 4: A Chrome extension built by a non-developer

How it worked before. Link building is one of the slowest parts of SEO. Every mention from another site is a negotiation by email, so our link-building specialist handles dozens of emails a day with hundreds of sites. The hard part was memory: we already work with this one, that one declined, this one needs a follow-up in a month. The list lived in a Google Doc, so every reply started with Ctrl+F. A small tool would fix it, and in most companies a small tool means a ticket to engineering that never reaches the top.

How it works now. He described what he wanted in plain words, in an ordinary chat, and sent screenshots whenever something broke. Claude wrote all the code. The result is a Chrome extension that highlights domains inside Gmail: green for "already have it," red for "not a fit," yellow for "promising." To tag a new site, he selects it and picks a color from the right-click menu. The database holds 800 domains, and each status shows the moment he opens an email.

Prompt you can adapt:

I'm not a developer. I want a Chrome extension that highlights website domains inside Gmail emails based on a list I keep. Each domain has one of three statuses: green = already working with us, red = not a fit, yellow = promising. When I select a domain in an email, I want to set its status from the right-click menu. The list should be saved in the browser so it survives a restart. Write the full code and give me step-by-step instructions for installing it in Chrome. I'll send screenshots if something doesn't work.

Before → after: 1–2 hours a day spent cross-checking a spreadsheet → zero. The status is in the email. He also says a long list of sites in one email no longer feels overwhelming.

A Chrome extension built by a non-developer

Use case 5: A morning digest as a Claude Cowork scheduled task

How it worked before. On any morning, a marketer's inbox holds partner replies, ad and analytics alerts, industry newsletters and requests from colleagues. Only a few need action today, but finding them means skimming everything. Email is also only part of it: follow-ups get promised in Slack threads and action items get agreed on calls. The real cost is the reply that never got sent.

How it works now. This is our simplest use case and the easiest one to copy. A Claude Cowork scheduled task runs every weekday morning and sorts the last 24 hours of email into categories the person defined. Connect Slack, and Claude also checks who is still waiting on you and what you promised. Connect a call notes tool such as Granola, and it adds the action items you agreed on. You start the day with one list. For more ways to use Cowork beyond this, this HackerNoon collection of Cowork tips is a good companion.

Prompt you can adapt (set it to run every weekday morning):

Every weekday at 9:00, go through my Gmail from the last 24 hours and group the emails into: 1) needs my reply today, 2) worth forwarding to a colleague (suggest who), 3) AI and marketing industry news from newsletters, one line each, 4) everything else, as a count only. Then check Slack for messages from yesterday where someone is waiting for my answer or where I promised to do something. End with one combined to-do list for today. Don't reply to or send anything yourself.

Before → after: the first part of the day went to sorting the inbox → one list, ready before you open email. The last line of the prompt matters: the task reads, it never sends.

A morning digest as a Claude Cowork scheduled task

Where it ended up

Across the five use cases, work that took weeks or hours now takes hours or minutes.

Task

Before

With Claude

What changed

Case study

1–2 weeks

~2 hours

Project team isn't interrupted

Custom sales page

Weeks

2–3 hours

Developers review, don't build

Blog post published

2–3 hours

~30 minutes

No platform workarounds

AI-search question list

An afternoon

Minutes

No spreadsheet exports

Link-building status check

1–2 hours a day

In the email

No engineering ticket

Morning triage

First part of the day

One list at 9:00

Nothing slips between tools

The rules we keep, and where the teammate model breaks

A teammate still gets reviewed. These rules sit in our project instructions:

  1. Review every output. We proofread every case study and article by hand. The less context Claude gets, the more editing you'll need, and unread output is how AI slop ships.
  2. No invented numbers. Figures come only from named sources: Jira, call recordings, approved case studies, Ahrefs. A missing number is written as [missing].
  3. A person decides what goes public. Claude sorts quotes into safe and confidential; a person makes the final call.
  4. Nothing is sent or published without approval. Scheduled tasks read and sort; they don't reply.
  5. Fix the instructions, not just the result. When Claude gets something wrong, we explain the mistake, then add the rule to the project so it applies to every new chat.
  6. Pick the simplest setup that works. A scheduled task was enough for the digest; link building needed a browser extension.

The model breaks in four situations: a connector doesn't allow the action you need, Claude gets too little context, nobody reviews the output, or the task is really a customer-facing product. The first three have workarounds. The fourth needs a different kind of build: an assistant that answers customers from a live ERP needs architecture, integrations, response validation and security. I wrote about that jump in what it takes to make a Claude prototype enterprise-ready.

Ground rules for ClaudeHow to set this up in your first 60 days

Start with one routine task your team does every week the same way: a report, a status update, a case study, a round of outreach emails. Then follow the order we use:

  1. Connect the systems where your context lives: task tracker, wiki, messenger, call recordings, email and CRM.
  2. Create one shared project for one team or one recurring task, and let Claude draft its instructions.
  3. Write down the boundaries: where numbers come from, what Claude can't change without confirmation, who approves the output.
  4. Start with 2–3 real use cases, then add one a week.
  5. Turn anything done twice into a playbook, and stable playbooks into skills.
  6. Show results to other teams. Whatever works for several teams becomes a company standard.

Anthropic's own guide to scaling workflows with Claude Cowork suggests a similar model, where a few champions build starter skills for their team.

Before you call Claude a teammate, check:

  1. Does the task repeat at least weekly, in the same structure?
  2. Does its data already live in systems Claude can reach through connectors?
  3. Are the instructions shared, or still in one person's chat history?
  4. Is every number traceable to a named source?
  5. Is one named person responsible for reviewing the output?
  6. Is a mistake cheap to catch before anyone outside the team sees it?
  7. Is it internal work, not a customer-facing product?

If most answers are yes, you have a teammate job. If the last one is no, you have a build.

FAQ

What does "Claude as a teammate" mean? It means Claude works inside a team's real systems under shared rules, instead of in one person's private chat. It takes three things: instructions that set Claude's role and rules, connectors that give it context from tools like Jira, Slack and Google Drive, and playbooks for recurring tasks.

How do marketing teams use Claude Cowork? Common Claude Cowork use cases for marketing are case studies drafted from project records, research and drafts for AI-search articles, website pages through Claude Design and Claude Code, small internal tools, and scheduled morning digests of email and Slack. A person reviews every output before it is published or sent.

What are Claude marketing skills? A skill is a reusable set of instructions that produces a fixed kind of output, such as a researched article outline or a monthly report. Marketing teams turn their most repeated playbooks into skills, so every run follows the same steps and rules.

Is it safe to connect Claude to HubSpot, Gmail and Slack? Official connectors give Claude controlled access to the data it needs, which is safer than pasting data into chats. On Claude Team and Enterprise plans, only organization Owners can add custom connectors. Add written rules on top: numbers only from named sources, and nothing sent until a person approves it. Connecting Claude to enterprise data goes deeper on MCP.

Can a marketing team do this without developers? Mostly, yes. Case studies, AI-search articles and the daily digest run on connectors and prompts, and the Chrome extension was built by someone who doesn't write code. Website pages are the exception: Claude Code writes the page, but someone technical should review the code and deployment before it goes live.


This article was published under HackerNoon's Business Blogging program.