> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI Bayesics #6: Claude Cowork - URL: https://bayesiantheory.com/ai-bayesics-6-claude-cowork/ - Published: 2026-04-24T11:53:22.000Z - Updated: 2026-10-05T14:45:19.000Z - Description: Harness the power of the SaaS-killer - Author: Gregg Bayes-Brown - Tags: AI Bayesics, artificial intelligence, Writing, Skills, Training, AI, Coaching, Writing with AI, AI training, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 In February 2026, over $800bn was wiped off global tech and software stocks. The reason: Claude Cowork launched. All very well and good, but what is it and how do you use it to your advantage? By the time you’ve finished this blog, you’ll be able to answer those questions. I’ve been using it since launch, and it has rapidly become the bedrock of my business. I can pass off work reliably to it, I’ve built a whole bunch of tools and apps which empower me during the day, I can schedule tasks, offload my finances, and pretty much anything else I imagine. The crucial bit here: you don’t need to be hugely technical or know code. There are, however, things you need to learn to get the most out of it. Before we begin: a quick note to regular readers. Yes, I had planned to do a Bayesics on how to research. But my one paying subscriber told me to cut to the chase, and I am a slave to my audience. We’re jumping ahead here, but I’ve done what I can to explain any terminology as best I can. Let’s get into it. [Subscribe now](#/portal/signup) ### What is Cowork The big shift in AI this year has been the move from agentic flows to agentic harnesses. 2025 agentic thinking went ‘how do we replace a human process with an autonomous machine?’ These are brittle, cannot interact with the human organisation around it, technical, cost jobs, and are prone to mistakes. In 2026, an agentic harness is as much about equipping the human as it is the model – a symbiotic partnership. Think of it like this: 2025 was a robot doing your job. 2026 is you in a mech suit. Anthropic’s Claude Cowork is a variant of that shift which makes agentic harnesses accessible to non-technical knowledge workers. I previously explained why this is a big deal when I covered models, apps and harnesses in detail in *My COO doesn’t sleep, eat or need a salary*. But a quick recap: - **Models.** GPT 5.4, Opus 4.7, Gemini 3.1 etc. The brains of the operation, if you will. If you’re not using a LocalLLM\[1\], you’ll either be accessing models via API\[2\] or… - **Apps.** [**Claude.ai**](http://claude.ai/?ref=bayesiantheory.com), Lovable, Copilot. These wrap around a model. Part of it will be the system prompt – the instructions the model receives as default. This why ChatGPT and Copilot can feel so wildly different despite both using GPT 5.4\. The app will likely also have specific tools or additional features – CustomGPTs, image gen, connectors\[3\], etc – which grant the model specific abilities. - **Harnesses.** Codex, Cowork, OpenClaw. As the name implies, these provide the model with a specific harness of agentic tools. These include the ability to use your computer, spin up sub-agents to perform concurrent tasks, and interact with other apps. So say we took the prompt ‘write me a daily briefing. Ask no questions. Do it now’. If we pasted into an API straight to Opus 4.7, Claude Chat on browser, and Claude Cowork on desktop, we’d get very different results despite the model being the same. The naked model knows nothing about you. Without a web search tool, it cannot access information beyond its cutoff date, which will be before today. The response you get will be generic and likely unhelpful and inaccurate. Claude Chat, on the other hand, can get up to date information. It also likely knows who you are and what interests you from memory. If you’ve set up instructions in the backend, it’ll know what tone and style to adopt. If you’ve used connectors to add your calendar and emails, it’ll scan those too, and give you your itinerary for the day. Claude Cowork goes further. It can see and work with your files and interact with your computer. It won’t just use a template for your briefing, but help you design one for all your work. It can help you set up a personalised dashboard for your briefing, and offer to set it to autonomously fire at 9 AM for you. You can assign it to do some of your tasks from that dashboard – desk research for a project you’re starting, an end of the month report, pre-meeting prep notes for the day ahead. It’ll then use skills you’ve built together – writing with your style guide, firing up a slide deck making skill, following a specific workflow, completing error and slop sweeps, rescheduling some time to review output in your calendar – to complete the work. You can also download additional skills and plugins from creators (such as Bayes Ltd) to expand what it is capable of, all with the ability to customise to your organisation and style. Like the harness which created it, agentic coding platform Claude Code, Cowork is capable of creating bespoke software, agentic flows, dashboards, AI powered procedures, apps and so on. ### How do I actually use it? Some assembly is required. First, you need a Claude subscription – free accounts aren’t going to cut it. You also need to leave your browser behind if that’s where you’re still interacting with AI. Go to [**claude.ai/download**](http://claude.ai/download?ref=bayesiantheory.com) and grab the desktop app. Install, open, and then you’ll find a tab at the top left-hand side which switches between modes – Chat, Cowork, and Code. Or just shortcut to Cowork with Ctrl-2. Alongside prompts, Cowork makes use of slash commands to invoke skills. The first one we’re going to use here is: > /setup-cowork This is the onboarding skill. It will walk you through the basics – teaching you about skills, plugins and connectors. Run through this to get the foundations in. ## Giving it a home Cowork creates a virtual machine on your computer but still needs a place where you can both interact. Open up File Explorer (click the search option next to the start menu and type file explorer) then navigate to “C:\\Users\\\[YOUR USER NAME\]\\” – this will be the folder which hosts your desktop, documents, downloads etc. Create a new folder – doesn’t matter what you call it, but let’s call it ‘cowork’ for now. Back in Cowork, select the folder drop down to the bottom left of the prompting window, select your folder, and star it: ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/0cfe459c-5822-4aa5-beb7-3ffd7158ff3d_752x582.png) This sets your folder as default. You may also note you can select Projects from regular Claude Chat, a feature which comes across into Cowork. These allow you to group chats and files around a particular topic – these can then be added to Cowork to provide more context to a task you’re working on. Back in your folder, you’re going to want to add some subfolders: About, Outputs, Templates, & Examples. You’ll also want some bespoke to you. Bayes, for example, has a clients, builds, admin folder and more. - About: add your brand instructions, your writing style guide, an overview of you personally and your organisation, and a [**Claude.md**](http://claude.md/?ref=bayesiantheory.com)\[i\]. If you don’t have any of these, ask Cowork to help you create them. - Templates: documents you’ll want to replicate often. As you refine your setup, you’ll likely want to add skills which explain the process to Cowork. Again, Cowork can help you make these if you ask. - Outputs: this is where Cowork will put its work. When you’re writing the [**Claude.md**](http://claude.md/?ref=bayesiantheory.com), you can add “About and templates are read only, all your outputs should go into /outputs” – this will help prevent Cowork from messing with core files and sources of truth. - Examples: finished examples of intended outputs. For example, you want to do a press release. Templates has the structure. Examples has the end result to emulate. So I’d have a folder in examples called ‘announcements’ and add a handful of the most impactful ones I’ve written to date. Finally, we’re going to add system instructions (see AI Bayesics #1). For Cowork, these are in settings under the Cowork tab. As before, if you’re unsure what to write, we can work with Claude here. > “I want to write global instructions for Cowork to define how we work together. Work with me back and forth. Interview me one question at a time, then let’s draft the instructions.” A good starting point: “read the about folder before starting any task”. ### Get connected Cowork thrives on connectors. These are MCP connectors which enable Cowork to directly work with other apps and your desktop. Some you’ll likely want to setup right away include your emails, your calendar/to-dos, and any apps you use religiously. To do this, hit ‘customise’ in the lefthand sidebar (it’ll be underneath ‘Dispatch’), click connect your apps in the next screen, then hook up anything you use regularly. ### What’s in it for me? If you’ve been following AI Bayesics from the start, you should already know how to answer this question. For everyone else, go to Cowork once you’ve done the setup, and write: > “I want to discover efficiency and effectiveness use cases for Cowork in my work. Read my folder and ask me any questions about my work where you need more information. Then, give me a list of all the use cases relevant to me.” See something you like? > “How do we setup X?” ### Things to know ### Continuous Improvement Cowork is insanely powerful, but the difference maker is what you bring to the dance. Once you’ve got it dialled in, Cowork makes few mistakes (if any – and the mistakes it does make are usually a result of what you’ve fed it). That said, you’ll still want to read through what comes out of it. As it is your work and in your domain, areas that need refinement, tweaks or corrections should jump out. If you’re finding something is repeatedly not adding up, it’s in your approach. For example, keep getting Americanisms when you’re using UK English? That needs to be an instruction. Cowork keeps missing a step in a process? That process needs to be defined as a skill, and Cowork needs to know to fire that skill (again, instructions). ### Use Markdown Markdown is a variant of text files which are both machine and human readable, making them ideal for context. They are also incredibly efficient – a PDF will burn 20x the tokens that the same content in markdown uses. Using them religiously will help keep costs down and avoid you blowing context every other prompt. You can make these in notepad, or simply ask Cowork to output something as a markdown. You can also take PDFs, word docs, etc and ask Cowork to turn them into a markdown file you can use as context. ### Use fresh sessions When you prompt any AI, it will read the entire conversation each time you do. Every turn of conversation adds another layer of context, which results in higher token burn. Additionally, as your context window fills up, the outputs tend to drop in quality. To get around this, avoid continuing in the same chat any longer than you have to. Each task should get its own chat. If you’re working on something bigger (like developing an app), break up the stages. For example, do your brainstorming, wrap up, then do research in a new chat, wrap up, then onto planning, stop, then execution. To give yourself a headstart in the new chat, you can finish your current session with: > “Save the key information from this chat as a research.md/plan.md/brainstorm.md etc and give me a prompt to setup the next stage.” ### Features - **Dispatch.** Your remote control for Cowork, allowing you to complete tasks on the go (so long as your PC is on somewhere). Incredibly powerful. - **Live artifacts.** New. I haven’t moved any of my tools onto this yet, but I can see where it’s going. A live artifact is a self-contained page — a dashboard, tracker, explorer, interactive report — that lives in your sidebar, persists across sessions, and can call your connectors for fresh data every time you re-open it. - **Scheduled tasks.** Claude fires a task at a time you pick. Daily briefing at 9am. Weekly pipeline check Monday morning. Monthly report on the last day of the month. Runs in the background, delivers the output. ### And finally… With all of this, you should now be setup to forge your own path with Cowork from here. I do, however, appreciate this may be something of a step up for many. I’ve been half thinking about doing a workshop online in the next week or so for people who would appreciate me showing them the ropes – drop a comment if this sounds useful, and I’ll set something up. Next week, I’ll dive into GPT 5.5 and OpenAI’s recent developments, which looks amazing. And maybe the research piece, if **Fionnuala O’Conor** lets me. \--- Thanks for reading. If this guide has proven useful, please do feel free to share with your contacts. A like, comment and share massively helps me build on what I’m doing here, so please do let me know you’ve been here! Also, please do consider becoming a subscriber (either paying or free) to my Substack as it genuinely makes a massive difference my end. --- \[1\]A locally hosted Generative AI model, such as Gemini’s Gemma 4, Qwen 3.5, or any number of peers. These use on-board GPUs, APUs and RAM rather than a model in the cloud, such as Opus 4.7. \[2\]Application Programming Interface, a set of protocols which enable applications to interact with each other. \[3\]A bridge that lets an app or model access another service. In AI tools, connectors wire a model into apps like Gmail, Calendar, or Slack so it can read from or write to them. --- \[i\] This is a core file for any Claude setup. It includes dos and do nots, common pitfalls to avoid, core instructions, and fundamental processes. The prompt you need is “I want to write a [**claude.md**](http://claude.md/?ref=bayesiantheory.com) to define our work together. Work with me back and forth. Interview me one question at a time, then let’s draft a [**claude.md**](http://claude.md/?ref=bayesiantheory.com)”. ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/46c7d674-1ddd-4496-908f-c85f74c91e8f_682x1500.png)