7 Steps to Build AI Workflows That Remember Your Business, Follow Your Process, and Save You Hours Every Week
This is the head start most people won't take until it's too late.
You’ve been chatting with AI for two years. Asking questions. Getting answers. Copy-pasting things into documents.
And every single time, you start from zero.
The AI knows nothing about you, your business, your audience, or how you like things done. You explain it all again. You paste in the same information. You describe your tone of voice for the hundredth time.
That era is over.
Anthropic’s Claude Cowork has changed how this works. It doesn’t just talk to you like it’s 2024. It takes action. It opens your files. It connects to your software. It follows multi-step processes you define, from start to finish, while you sit back and relax with a hot drink.
And it remembers!
It builds context about you over time. It can access business knowledge on demand through files and skills you set up once and reuse forever.
The feature that ties all this together?
It’s called a “skill.”
If your job involves repeating similar workflows, then this will be a game-changer.
Because the gap between “people who chat with AI” and “people who build context-rich AI workflows” is about to become the biggest productivity divide in the modern workplace.
This guide walks you through the entire setup in 7 steps.
By the end, you’ll know how to give Claude Cowork deep context about your work, connect it to the tools you already use, and build custom skills that automate your most repetitive tasks. With output that actually sounds like you.
How Claude Cowork fits alongside Claude and Claude Code
It will help you to understand where Cowork sits in the line-up before we get into it.
Claude (the chat interface at claude.ai) is what most people already know. It’s great for brainstorming, Q&A, and working through ideas in conversation.
Think of it as your thinking partner. Claude also has a memory feature that learns your preferences over time.
Things like your role, your communication style, and topics you return to often. This memory carries across conversations, so Claude gets better at anticipating what you need the more you use it.
Claude Code is a command-line tool built for developers. It’s designed for building production-ready applications, writing and debugging code at scale.
Claude Cowork is the new middle ground. And it’s the one that matters most for day-to-day office work. It runs as a desktop app, accesses your local files, connects to your software stack, and executes multi-step workflows through ‘skills’.
It also inherits the planning capability from Claude Code. That means it breaks tasks into steps and works through them methodically.
The way this will likely shake out: Cowork for daily tasks and workflows. Claude Code for building software. Claude chat for brainstorming, analysing, summarising and drafting content.
The context principle (why most people get average AI output)
Before we get into the setup, there’s a concept behind everything in this guide. It’s the single biggest reason some people get great results from AI while others get generic, forgettable output.
Context.
The quality of what an AI produces is directly tied to how much it knows about you, your business, your audience, and the specific task at hand.
Most people give AI almost no context. They type a one-line prompt and expect magic. What they get is a bland, one-size-fits-all response that isn’t very helpful. Then they walk away thinking that ‘AI’ is a bit crap.
Now think about how you’d brief a new starter. You wouldn’t just say “write me a newsletter.” You’d share as much information as possible to get the output you want.
You’d show them past outputs that worked. You’d talk about your clients, their brand, previous projects. You’d walk them through the process step by step until they understood what good looks like.
That’s exactly what Claude Cowork lets you do. But permanently.
You set up the context once. You save it in files, in memory, and in skills. Every future task starts from a position of deep understanding. Not from zero.
This is the bit that changes everything. The people getting outstanding AI output aren’t better prompt engineers. They’ve built better context systems.
Here’s how the layers of context stack up:
Memory
Learns your preferences, your role, your recurring topics, your projects, your colleagues and your communication style across conversations.
It builds over time without you needing to do anything specific. You can also tell Claude to remember things directly: “I always write in UK English,” “My audience are brand managers who work in large pharma companies,” “I prefer short paragraphs and direct language.” This memory sticks and shapes every future response.
Business context folders
The files you create once and point Cowork at whenever you start a task. Your client brand documents, previous projects, the style guide, content strategy - whatever you like. These give Cowork specific knowledge about your business that memory alone can’t capture.
Skills
Combine context with process. They tell Cowork what to do, how to do it, what files to reference, and when to ask for your input. Skills are where context becomes workflow.
When you layer all three together, the AI stops producing generic content. It produces work that sounds like you briefed a colleague who’s been at your company for years.
Every step in this guide builds on this principle. Keep it in mind as you set things up.
Here’s what we’ll cover:
1. Give Cowork access to your files
2. Connect it to your software stack
3. Understand why skills change everything
4. Use pre-built skills to start fast
5. Build your first custom skill
6. Convert your existing Claude Projects and GPTs into skills
7. Use code execution for data tasks
Step 1: Give Claude Cowork access to your actual files
Most people still copy-paste text into a chat window. That works for quick questions. It falls apart the moment you need to work with real documents, folders, or data sitting on your machine.
Claude Cowork lets you point directly at a folder on your computer. Your project folder. A folder full of briefs. It reads everything inside, understands what’s there, and takes action on it.
You can create dedicated folders that hold everything Cowork needs to know about your business and the projects you’re working on.
When you start a new task, you point Cowork at those folders. It reads your context and applies it automatically. This becomes especially powerful once you start building skills (more on that in Step 3).
Tips to get this right
Create a “Project” folder with 3-5 key documents: your scope of work, timelines, client brand, briefs, and any strategic docs you reference often. This folder becomes the foundation for everything you build later.
The more specific and detailed these documents are, the better Cowork’s output will be. A one-page persona of your target audience is good. A three-page ICP with real customer quotes, common objections, and specific language your audience uses is far better.
Also create an “examples” folder. Fill it with past work. Website content, Newsletters that performed well. Social posts that got engagement. Reports that landed with stakeholders. Cowork can study these examples and match your style with surprising accuracy. Examples teach AI things that instructions alone cannot.
Cowork will ask you clarifying questions before it acts. This planning step is inherited from Claude Code’s architecture. It’s one of the features that makes Cowork feel more reliable than a standard chatbot. Don’t skip past the questions. The better your answers, the better the output.
You need a Claude Pro, Team, or Enterprise subscription. Cowork isn’t available on the free tier. It only runs on the Claude desktop app (not in the browser).
Step 2: Connect Claude Cowork to your software stack
There are three ways to make this connection happen.
Built-in connectors
These are the easiest. Open Claude’s settings, head to the Connectors panel, and you’ll see a list of supported software.
Toggle one on, authorise the connection, and Cowork can immediately start pulling data and taking actions inside that tool.
If your tool is listed, you’re up and running in under a minute.
MCP servers
These are easier to set up than it sounds. And it’s not a web server.
They handle the tools that aren’t built in. MCP (Model Context Protocol) is the standard Anthropic uses to connect Claude to external software.
Most major tools now publish their own MCP setup instructions.
If there isn’t a ready-made plugin, you search for your tool’s MCP documentation, copy a JSON config block, paste it into the Claude desktop config file (found in the Developer section of settings), and save.
It sounds technical but it’s usually just copying a block of text into the right place. Automation platforms like n8n can also create custom MCP servers for any tool with an API.
That means you can connect nearly anything to Cowork even if it doesn’t have official MCP support yet.
Browser use
This is your fallback. If your software doesn’t have a built-in connector or an MCP server, Cowork can open a browser and interact with the tool directly.
Clicking buttons and reading screens the way you would. It’s slower than an API connection but it works with nearly anything.
One nice detail: browser tasks can run in the background inside Cowork while you work on a separate task in a new window.
Each connected tool also becomes another source of context. Cowork can pull your existing tasks, project notes, pipeline data, and meeting records into a workflow.
That gives it real-time information about what you’re working on right now.
Tips to get this right
Start with the tool you use most. For many people, that’s Notion, Asana, or a similar project management tool.
Connect that first and test a simple query like “show me my open tasks” before connecting anything else.
If Cowork doesn’t have a direct connection to a tool and you ask it to access that tool, it will often trigger browser use automatically. You don’t always need to tell it explicitly.
Browser-based research tasks are a good early test. Ask Cowork to research a topic on a specific platform. It’ll browse, scroll, and compile findings while you carry on with other work.
Step 3: Understand why skills change everything
This is where you go ‘next-level’ in your AI knowledge.
Think about the tasks you do every week. Creating briefs, analysing data, writing a scope of work or a requirements document. Reviewing a content brief. Packaging a video idea into a title and thumbnail. Prepping slides for a presentation.
Each of these tasks follows a specific process. You probably have a mental checklist. A preferred order of how you do it. A set of reference materials you pull from each time.
Now imagine you could save that entire process as a single reusable package. The steps, the context files, and the instructions. All in one place.
That’s a ‘skill’.
It’s is a saved set of instructions, a defined process, and a collection of knowledge sources. Together they tell Claude Cowork exactly how to execute a specific workflow. Think of it as a system prompt that actually does things.
Skills are where your context (business docs, examples, preferences) and your process (steps, decision points, actions) come together into something repeatable.
What makes skills different from Claude Projects or custom GPTs?
You can trigger multiple skills in the same conversation. Writing a project? Trigger your “client brief to WBS generator” skill, then your “scope of work” skill, then your “project plan” skill. All in one session.
With Claude Projects or custom GPTs, you’d be jumping between three separate interfaces.
Skills only load their instructions and knowledge sources when triggered. This keeps the context window clean. You’re not overloading Cowork with irrelevant information from five different workflows when you only need one. This is a smarter approach than stuffing everything into a single system prompt and hoping the AI figures out what’s relevant.
Skills can also include instructions to update external software at the end of a workflow.
After you finalise a document, the skill can update your Notion pipeline or project tracker automatically through the connectors you set up in Step 2.
Why this matters more than full automation
A lot of daily work needs a human in the loop. Project planning, content creation, strategic decisions, client communications. These tasks follow a process, but they require judgment at multiple points.
A fully automated workflow on a platform like n8n or Make can feel rigid for this type of work. You end up fighting the automation or hopping between interfaces when the task has nuance.
Skills sit in a sweet spot. They automate the repeatable parts while keeping you in control at the decision points.
Step 4: Use pre-built skills to start fast
You don’t need to build everything from scratch.
Claude Cowork comes with built-in skills. And a growing community is sharing skills publicly.
Built-in skills are found in Claude’s settings under Capabilities.
Scroll down to “Example Skills” and you’ll see what Anthropic has included out of the box. Canvas design, MCP builder, a skill creator (which helps you build new skills), and others.
Trigger any of them by telling Cowork to load the skill by name. It will read the skill’s instructions, ask you the right questions, build a plan, and execute step by step.
Community skills are where it gets interesting.
Three marketplaces have popped up as the main hubs: smithery.ai/skills, skillhub.com, and skillsmcp.com. Between them, there are thousands of user-created skills covering everything from ad copywriting and SEO analysis to financial modelling and code review.
Some of them are surprisingly good.
Tips to get this right
Download a community skill and test it before building your own. This gives you a feel for how skills are structured and what good instructions look like.
Most skills come as a zip file. Upload it through Claude’s settings under Capabilities and it appears in your skill library straight away.
Community skills won’t have context about your specific business. This is where layering context makes a massive difference. Pair a community skill with your “business context” folder from Step 1.
Give Cowork access to your ICP, voice guide, and example work alongside the skill.
A generic “ad copywriting” skill paired with your specific business context will produce output that’s miles ahead of either one used alone.
Step 5: Build your first custom skill (the walkthrough method)
This is the most powerful approach.
Here’s the idea. Walk through one of your regular tasks with Claude Cowork, step by step, doing it manually just once. At the end, ask Cowork to save that entire process as a skill.
Next time, you trigger the skill and it follows the same process automatically. All the context baked in. Boom!
Let’s help a colleague out with this one….
Say they repurpose video content into newsletters every week. The process might look like this: get the video transcript, brainstorm subject lines, pick one, write a hook, pick one, then draft the full newsletter in a brand voice.
They can walk through this once with Cowork. At each step, tell it what to do and give it the reference materials it needs. The voice guide, newsletter examples, content strategy etc.
Have Cowork use a browser to find a video and download the transcript for you.
Each step is something they’d normally do manually. But now Cowork is learning the process as you go.
When you’re done, you say: “Save this as a skill.”
Cowork packages the entire workflow. Every step, every reference file, every decision point where it should pause and ask for your input.
Tips to get this right
Don’t try to automate a complex workflow on your first attempt. Pick a task that has 3-5 clear steps.
Be specific about the order of operations during your walkthrough. If you always brainstorm subject lines before writing hooks, make that explicit. The skill will follow whatever sequence you demonstrate.
Include your reference files during the walkthrough. The skill will remember to load them each time it runs. This is the moment where all the context documents you created in Step 1 start paying off.
After the skill is saved, test it straight away with fresh input. You’ll spot gaps in the instructions quickly. You can update the skill on the fly. Pay attention to whether the output reflects your voice and context. If it doesn’t, the fix is usually adding more specific examples or tightening the context files. Not rewriting the skill instructions.
Step 6: Convert your existing Claude Projects and custom GPTs into skills
If you’ve already built Claude Projects, custom GPTs, or system prompts in other tools, you’re sitting on a goldmine.
All the context and process knowledge embedded in those setups can be migrated into Cowork skills.
Take your existing system prompt. Grab the knowledge sources and reference files attached to it. Hand everything to Cowork and say: “Create a skill out of this.”
Cowork will package your existing setup into a skill format. All those prompts and processes you’ve refined over months of trial and error are now portable, combinable, and reusable inside a single Cowork session.
This is the fastest way to build a library of skills and (if you’re like me) you’ve been using AI seriously and probably have 5-10 custom setups scattered across Claude Projects, ChatGPT, and other tools.
Each one holds accumulated context. Your refined instructions, your curated examples, your specific preferences for that type of task.
Converting them into Cowork skills means you stop jumping between interfaces. Everything lives in one place. You can trigger any combination of skills in the same conversation.
Tips to get this right
Copy the full system prompt from your existing Claude Project or custom GPT. Don’t paraphrase it. Cowork will use the exact instructions to build the skill.
Include all knowledge sources: brand guides, example outputs, process documents, templates. The skill will reference these each time it runs. These context files are the reason your old setups worked well. Without them, the skill is just instructions without understanding.
After conversion, add instructions for any software updates you want the skill to trigger through your connected tools. “After the title is finalised, update my Notion pipeline with the chosen title and move the status to ‘In Progress.’” This is something your old Claude Project couldn’t do.
You can combine multiple old setups into a single skill. Or keep them separate and trigger them in sequence. Both approaches work. Keeping them separate gives you more flexibility to mix and match.
Step 7: Use code execution for data tasks
Claude Cowork can also run code. This opens up a whole category of tasks that conversation alone can’t handle.
This is different from Claude Code which is primarily for building applications. Code execution in Cowork is for getting specific tasks done. Data visualisation. File processing. Formatting.
The most practical use case is data visualisation. Point Cowork at a spreadsheet or CSV file and ask it to create charts, graphs, or summary tables. It writes and runs the code behind the scenes. You get the visual output without touching a code editor.
Image formatting is another common one. Need to resize a batch of images, change aspect ratios, or convert file formats?
Cowork handles it through code execution inside the same workspace where you’re doing everything else.
Tips to get this right
Think of code execution as code in service of a task. Not building an application. Use it for data work, file processing, and formatting. For anything more complex, Claude Code is the right tool.
Be specific about what you want to see. “Create a bar chart showing views per video from this spreadsheet” gets better results than “analyse this data.”
You can combine code execution with skills. A reporting skill could pull data from a connected tool, run analysis code, and output formatted charts. All in one workflow.
Your next move
The shift from “chatting with Claude” to “working with Claude” happens the moment you build your first skill in Cowork.
Everything before that is just a conversation. Everything after is a workflow.
But the real win isn’t the tool. It’s the context you feed it.
The people getting the best results from AI right now aren’t the ones with the cleverest prompts.
They’re the ones who invested an afternoon building context. Filling project folders with examples of real ‘work’ - whatever that looks like. Stuff that Claude can reference on demand.
Here’s how to get started.
Spend 30 minutes creating your context folder.
Then pick one task you do every week that follows a clear, repeatable process. Walk through it once with Claude Cowork, step by step. Save it as a skill. Run it again with fresh input.
That first skill will save you time every single week going forward.
And once you see the difference context makes to output quality, you’ll find yourself converting every Gemini Gem, every custom GPT, and every repeatable task in your day into a Cowork skill.
Figure this out now and you will have a compounding advantage. Every skill you build makes the next one faster.
Every piece of context you add makes the output sharper. Every workflow you automate frees up time to think about the work that actually requires your brain.
Remember: You + AI = Superpowers.
Cheesy, but true!