Claude Cowork: A practical guide for product marketers [Claude series 2 of 6]

In part 1 of this series, I mapped the full Claude ecosystem – Chat, Projects, Cowork, Skills, Connectors, Code. If you missed it, start there.

This edition goes deep on the one that changes PMM work the most: Cowork.

I’ve been pressure-testing it across real workflows for months.

What follows is everything I’ve learned – the setup, the workflows, the gotchas, and the honest limitations.

Let’s begin.

First: understand what you’re actually running

This matters because it changes how you use it.

When you open the Cowork tab in the Claude desktop app, you’re not switching to a different chat mode. You’re launching a sandboxed Linux VM on your machine. Inside that environment, Claude can write and execute code, create and manage files, run scripts, and – with the Chrome extension installed – control your browser.

It doesn’t work linearly. When you give Cowork a complex task, it decomposes it and spins up parallel sub-agents.

Ask it to research five competitors and build a comparison deck: it might spawn five agents simultaneously – one per competitor – then synthesise everything into the final output. This is why Cowork can handle tasks that run for 10, 20, even 30+ minutes. It’s orchestrating, not just responding.

And unlike Chat, which wraps up after a few exchanges, Cowork keeps going.

It runs for extended periods as long as your desktop app is open and your machine is awake.

Start a task, go do something else, and come back to the finished files in your folder.

The mental model shift: Chat is a conversation. Cowork is a brief.

You describe the end state, provide the inputs, set the constraints, and delegate.

Claude plans the work, executes it, and delivers output as real files – .docx, .pptx, .xlsx, .pdf – saved to your computer.

Setting it up properly (most people don’t)

The number one reason people try Cowork once and go back to Chat? They skipped the setup.

They typed a cold prompt with no context and got a generic result. That’s not a Cowork problem. That’s a setup problem.

Here’s what to do once. It takes about 15 minutes, and it changes every session after.

1) Install and configure

Download the desktop app from claude.com/download. You need a Pro plan ($20/month). Open the app, click the Cowork tab. Select Opus 4.6 as your model. Turn on Extended Thinking. These are your defaults.

2) Create your working folder

Cowork needs a dedicated folder on your computer to read from and write to. Create one inside a cloud storage folder (Dropbox, Google Drive, iCloud) – this way your files are local (which Cowork requires) but also synced and backed up. Cowork can read, create, edit, move, rename, and delete files in this folder. This is real file management on your actual machine.

Inside that folder, set up a clean structure:

• /context – your about-me, voice, and rules files (more on these below)
• /inputs – raw materials: transcripts, research, briefs, data
• /drafts – work in progress
• /final – approved deliverables
• /archive – old versions (never delete, always move here)

Add a _INSTRUCTIONS.md file in your main folder (not inside any subfolder) with your naming conventions and folder rules. This is a Markdown file, which is just a lightweight text format you can create in any text editor. Claude reads it automatically every session and follows it without you having to repeat yourself.

For example, mine says: always save new files to /drafts first, use the naming format YYYY-MM-DD-project-description.docx, never overwrite existing files (create a new version instead), and never delete anything (move to /archive)

3) Build your context files

This is the single most important setup step. Create three .md files in your /context folder:

• about-me.md: Who you are, what you do, what you’re working on. Your role, company, product, market, current priorities. Be specific: “I’m the PMM lead for a B2B data platform targeting mid-market RevOps teams, currently preparing for a Q3 launch of our analytics module” gives Claude vastly more to work with than “I’m a product marketer.” This is mine, for reference:

• voice-and-style.md: How your brand communicates. Tone descriptors, formatting preferences, phrases you use, phrases that sound wrong. Include examples of copy you’re proud of.
• working-rules.md: How you want Claude to operate. Ask clarifying questions before starting? Default to .docx? Never overwrite existing files? Create new versions instead of editing originals? Keep outputs under a certain word count?

This is what it will look like in your folder:

A quick note on structure: three to four focused files is the sweet spot. You might be tempted to consolidate everything into one massive file, or split into ten micro-files.

Every word loaded into context costs Claude thinking room, so one bloated file means unnecessary noise every session. But scatter context across too many files, and important instructions get overlooked. Three focused files, each doing one job well, is the right balance.

Remember, these files are your leverage.

Claude reads them at the start of every session. You stop re-explaining who you are and start getting output that sounds like your team produced it.

4) Set global instructions

Go to Settings → Cowork → Global Instructions and paste your persistent preferences: E.g:

# RULES (never break these)1. Never overwrite existing files. Always append -v2, -v3, or a date to the filename.2. Default to Google Doc for documents, Google Slides for decks, Google Sheets for data. If you can’t create the native format, stop and tell me before falling back to .docx/.pptx/.xlsx.3. Always read all files in the selected folder before responding.4. Ask clarifying questions before starting work.5. Read voice-and-style.md (in /context) before writing anything. Match the voice exactly.# WORKFLOW- Before saving a file, check the destination folder for name conflicts.- When research is involved, show me your sources.

This will now apply to every session automatically. A note: if all your rules are packed into one paragraph, everything runs together, so it’s easy for Claude to grab the gist and miss a specific rule. That’s why a numbered list works best.

5) Connect your tools

Go to Settings → Connectors and enable the integrations that matter for PMM work: Slack (competitive intel, sales feedback, product updates), Gmail, Google Drive, HubSpot or your CRM, Google Calendar. You authenticate once per tool. After that, Claude pulls live data from them in every session.

6) Install Claude in Chrome

It’s what unlocks Cowork’s full capability. The Chrome extension lets Claude browse websites, extract data from dashboards, fill forms, screenshot pages, and navigate the open web on your behalf. Without it, Cowork can only work with local files and connectors. With it, Cowork can operate across any website you give it access to.

Install from the Chrome Web Store, log in with your Claude account. You get three levels of browser autonomy: ask before every action (safest – start here), always allow on specific trusted sites (good for tools you use daily like Google Analytics, LinkedIn, Substack), or act without asking (maximum speed).

Even at the highest autonomy level, Claude still prompts you before purchases, permanent deletions, or security changes. That guardrail can’t be turned off.

Where Cowork changes PMM work

Here’s where it gets concrete.

Competitive battle cards

Give Cowork a competitor’s URL, their recent product announcements, and your own positioning doc from /inputs. Cowork browses the competitor’s site, reads their messaging, cross-references it against your positioning, and produces a structured battle card – strengths, weaknesses, talk track, objection handling, landmines – saved as a .docx in your /drafts folder.

Let me show you what a good prompt looks like for this:

“Read my positioning doc in /inputs/positioning-v3.md. Browse [competitor.com] and their blog. Produce a 2-page competitive battle card structured as: Company Overview, Key Strengths, Key Weaknesses, Head-to-Head Comparison (feature, messaging, pricing), Sales Talk Track (3 talking points), Landmines to Avoid. Save as .docx in /drafts. Match the voice in my voice-and-style.md.”

Compare that to the Chat-style prompt most people write: “Help me create a battle card for [competitor].” The difference in output quality is enormous. Context, end state, constraints. That’s the formula.

Customer research synthesis

Drop 15 interview transcripts into /inputs. Ask Cowork to extract themes, pull verbatim quotes as evidence, identify patterns across interviews, and produce a structured research summary with themes ranked by frequency. What used to be a week of qualitative coding becomes a task you check back on.

Sales enablement decks

Upload your messaging framework and win/loss data. With Slack connected, Cowork can pull the most common objections from your #sales-feedback channel. It produces a .pptx – real slides with layouts, speaker notes, and objection-handling scripts.

Weekly competitive monitoring

This is where Cowork’s /schedule feature earns its keep. Set up a recurring task: every Monday, Cowork checks three competitor websites for product updates and pricing changes, scans their blog and press page, pulls any relevant news, and compiles a digest saved to your /final folder. This scheduled task persists across app restarts.

As long as your machine is awake and Claude Desktop is open when the schedule triggers, it runs. If your machine is asleep, that run is skipped (it doesn’t queue up.) Start with something simple and low-stakes like this before scheduling anything with real consequences.

Messaging testing prep

Hand Cowork three positioning options and ask it to generate stimulus materials for each – landing page hero copy, email subject lines with body variants, one-pager drafts. It saves each as a separate file. Ready for stakeholder review or testing without any reformatting.

Managing context across sessions

Here’s something most Cowork guides gloss over, and it matters for any PMM running a multi-week project.

Cowork has a context window – Claude’s working memory for the current session. It’s large, but not infinite. In long sessions, older details get compressed. You’ll notice it when Claude forgets something you mentioned earlier, re-asks a question, or responses start feeling more generic.

The fix is a practice I now use at the end of every substantial session: ask Claude to write a HANDOVER.md. This is a structured summary of everything: what’s been done, what worked, what didn’t, key decisions made, and what needs to happen next.

Written so that a fresh Claude session can read it and pick up exactly where you left off.

Start your next session with: “Read HANDOVER.md in /context and continue from where the last session left off.”

No more lost progress between sessions.

Skills and projects – a quick distinction (full deep-dive next week)

You’ll hear both terms a lot. They’re different things.

Skills are the how – reusable playbooks that teach Claude how to execute a specific type of task. A battle card skill knows the methodology, the structure, the format. A launch brief skill knows the sections, the frameworks, the quality bar. Skills work across every surface – Chat, Cowork, and Code.

Projects are the what – persistent context about what you’re working on. Your product, your market, your competitors, your launch timeline. Projects hold the information. Skills apply the expertise.

The best output comes from combining both. A battle card skill that knows the methodology + a project that knows your product and competitive landscape = output that would take a junior PMM a full day.

There’s also a distinction between Cloud Skills (synced across devices, always active) and Local Skills (.md files in your working folder, activated only in that folder). Cloud for universal PMM processes. Local for client-specific or project-specific work.

I’ll cover all of this: how to find skills, install them, build your own, and set up your PMM toolkit – in next week’s edition.

Security and permissions: be intentional

Cowork has real access to your machine. That’s what makes it powerful, and it’s also why you should be deliberate about boundaries.

Folder access: Only give Cowork access to your dedicated working folder. Never your entire computer. Keep sensitive files – credentials, financial records, personal documents – outside this folder entirely.

Browser access: Start with “ask before acting” for all sites. Promote trusted tools (Google Analytics, LinkedIn, your CMS) to “always allow” as you build confidence. Be cautious with unfamiliar sites – web content is a vector for prompt injection, where hidden instructions on a page could influence Claude’s behaviour.

Connectors and plugins: Only install what you actually use. Each connector expands what Claude can access. Each plugin expands what Claude can do. Stick to verified ones from the Claude directory. Review permissions before installing.

Scheduled tasks: Start with monitoring and reporting (low risk, easy to review). Check outputs after each run. Don’t automate anything that sends messages, makes purchases, or takes actions that are hard to undo – at least not until you’ve built confidence in the output quality.

This isn’t about being paranoid. It’s about being intentional with a tool that has real agency on your machine.

Honest limitations

Worth knowing upfront rather than discovering mid-project:

No memory across sessions: Claude doesn’t remember previous conversations. That’s what handover files and context files solve – but you have to build the habit.

Desktop only (for now): Cowork runs in the Claude Desktop app. It doesn’t work on mobile or web, and sessions don’t sync across devices. Dispatch (see below) is starting to change this.

Must stay awake: Your computer and Claude Desktop need to be running for tasks and scheduled automations to execute. Close the lid or quit the app, and active tasks stop. Scheduled tasks that fire while your machine is asleep are skipped entirely.

Sessions aren’t shareable: You can’t hand a Cowork session to a colleague. You can share the output files – but not the session itself.

These are real constraints. They don’t make Cowork less useful, but knowing them upfront means you work with the tool rather than fighting it.

Dispatch: Cowork from your phone

Anthropic recently launched Dispatch as a research preview within Cowork. It lets you send tasks to Cowork from your phone.

You text the instruction, Claude executes it on your desktop, and you come back to finished files in your folder.

Heading into a meeting and need a competitive brief ready when you’re out? Send it from your phone. On the train and want to kick off a research synthesis? Same thing.

The bottom line

Cowork is the layer where PMM work actually gets produced. Not brainstormed. Not drafted in a chat window. Produced, as real files, on your computer, in the format your stakeholders expect.

The setup takes 15 minutes. The context files take another 15. After that, every task you run benefits from the infrastructure you’ve built. And it compounds: every context file you refine, every folder rule you add, every handover you write makes the next session better.

And to make it all stick a little easier, here’s a cheat sheet you can save:

Next week: The complete product marketing Claude skills pack – Everything you need to master Claude skills for product marketing.

The full series: Mastering Claude as a Product Marketer

Part 1: Mastering Claude for product marketing – What everything actually is. The map.

Part 3: The complete product marketing Claude skills pack – Everything you need to master Claude skills for product marketing

Part 4: The Claude marketplace for product marketing – Skills, plugins, and connectors. How to find them, install them, and build your own PMM toolkit.

Part 5: Claude Code for product marketing – When and why you’d go here. And why the jump from Cowork is smaller than you think.

Part 6: Bringing it all together – Claude for product marketing – How Chat, Cowork, Code, Skills, and Connectors work as one system. The full PMM Claude setup, from first install to daily workflow.