I run Prosper entirely inside [Claude Code](https://claude.com/product/claude-code). The full structure is there: context files, skills, the whole production workflow. It works well. I am used to it and I enjoy it. I wasn’t looking to replace it. But [Cowork](https://docs.anthropic.com/en/docs/cowork)), Claude’s project workspace that keeps your context, skills, and tools loaded between sessions, is becoming more and more capable, and I wanted to test a specific question: could you get the same infrastructure without touching a terminal? People open Cowork, use it like Chat, and miss the entire point. I imported my entire Prosper setup into Cowork: context files, skills, connectors. And ran it for 3 weeks to find out. What I discovered: Cowork isn’t a simplified Claude Code. It’s the same operating system, accessible through a mobile app and direct import from Claude Chat projects. Here’s what that unlocks in practice: I was at a coffee shop when someone mentioned a new AI tool I hadn’t heard of. I opened Cowork on my phone, typed “fast-research \[tool name\]”, and triggered a Dispatch job. By the time I got back to my desk, the brief was waiting: 3 angles worth exploring, audience fit assessment, and a recommendation on whether to cover it. All from my phone. No terminal, no config files. (Yes, Claude Code now has [Channels](https://code.claude.com/docs/en/channels) that let you talk to Claude through Telegram while running a session. But setting up Dispatch in Cowork is significantly easier: no webhook configuration, no bot setup, no channel management. It just works.)

/fast-research running on mobile via Dispatch — showing the skill triggered from phone — /fast-research running on mobile via Dispatch

/fast-research running on mobile via Dispatch

That’s what you get when Cowork acts like infrastructure instead of a chat tool. And that’s what this post is about - how the six layers work together, and how to build your own. ***If you follow this through, you’ll have:*** 1. The OS infrastructure that makes every feature click into place (5 minutes to understand) 2. How to import your existing Claude Chat project into Cowork (30 seconds) 3. The exact documents to load as your personal context, with role-specific guidance 4. My personal examples of files, skills and prompts 5. A working understanding of Skills, Connectors, Scheduled Tasks, Dispatch, and how they interact 6. A setup you can build in an afternoon and run for months

------------------------------------------------------------------------

## Why Claude Cowork Is an AI Operating System

An OS has a few core components: A **personal** **context** - the persistent layer that loads every time. It knows your standards, your constraints, your examples. Without it, nothing else runs correctly. **Installed** **apps** - tools you’ve set up once that run with a single command. You don’t rebuild them every session. **Connectors** - integrations with external systems. Your OS talks to other things, not just itself. **Background** **processes** - tasks that run without you initiating them. The OS does work while you’re doing something else. **Dispatch** - the orchestration layer. When a job is too big for one process, dispatch breaks it into parallel tasks and coordinates the output. **Visual** **rendering** - the display layer. When the OS produces something you need to see and interact with: a planning board, a dashboard, a chart - it renders right there. You don’t open a separate app to view it. Cowork has all six. In my experience, people set up the first layer, then stop. And even that first layer rarely gets the attention it deserves. ## How to Import Your Claude Chat Project into Cowork

If you already have a project set up in Claude Chat, with instructions and knowledge files, you don’t need to rebuild it from scratch. Cowork has a direct import. Here’s exactly how I did it: *1. Open **Cowork** and click **Projects** in the left sidebar*

Cowork home screen with Projects tab selected

2. *Click* ***New*** ***Project***

3. *Click* ***Import a Project***

4\. It shows you a list of your existing Claude Chat projects, pick the one you want to bring over

That’s it. Your instructions and knowledge files come across automatically.

I asked Claude directly what files it had access to and it listed all 17 files from my Prosper project: past newsletter issues, reference docs, workflow tools. Everything was there.

Claude listing all 17 imported project files in Cowork, organized by category — Claude listing all imported project files in Cowork, organized by category

Claude listing all imported project files in Cowork, organized by category

The import didn’t take long. My Prosper project, with all its context files, was available in Cowork immediately. Share *A couple of things to note:* \- The Cowork Project that you will be working in will not affect the Project in Claude Chat. \- You’ll want to review the instructions after import. Some things that worked in Claude Chat’s Projects behave differently in Cowork’s folder structure. I spent about 20 minutes adjusting constraints that needed to be more explicit. Worth a pass before your first real session.

The “Import a Project” screen showing that changes in Cowork won’t affect projects in Chat

------------------------------------------------------------------------

## Personal Context and Instructions

Every Cowork folder can hold “.md” files that Claude reads automatically before any task. This is your personal context. It loads before the session starts. If you use Claude Code, you know “CLAUDE.md” the file that loads persistent rules into every session automatically. Cowork has the same thing, just surfaced differently. It’s called Instructions, and it lives in the top-right panel of your project.

Cowork interface showing the Instructions panel location

The mechanic is identical: whatever you put there is loaded before every session. No need to re-explain yourself. No need to paste a system prompt. It’s always there. The mistake I see most often: people either don’t have these files at all, or they write them as preferences (”I like concise writing, professional tone”) rather than constraints. A preference tells Claude nothing. A constraint tells it what to reject. (Dense context files also save tokens.

------------------------------------------------------------------------

I wrote about 16 specific ways to cut Claude token burn:

#### Save Claude Tokens: 16 Tactics for Chat, Code & Cowork

Ilia Karelin

Apr 3

Read full story

------------------------------------------------------------------------

Share My Prosper folder has several context files. I’m not going to share the exact content of all of them. They’re too specific to me. But here’s the pattern that makes them work, and what you’d build for your own work. **The rule that matters most:** your context files should be mostly negative space. Not what you want. What you’d reject.

``` shiki “I’m responsive to client needs” ```

Tells Claude nothing. Everyone thinks they’re responsive.

``` shiki “I never confirm a timeline without naming the specific deliverable, the decision I need from them, and what happens if they don’t respond by Friday.” ```

Specific enough to apply. Here’s how the document categories map to different types of work: **Newsletter writers** need voice rules (what you’d never write), audience profiles (what they’re tired of reading), format defaults, editorial positions. **Consultants and agencies** need communication standards, delivery criteria, non-negotiables, business context in your own words. **Content** **creators** need content pillars, script formats, audience response patterns, what you’d never make. **Analysts and researchers** need output formats, rigor standards, how you structure findings, what vague language you reject. **Business** **owners** need business context, recurring task formats, client communication style, your tone at its best. The files don’t need to be long. Mine average about 150-200 lines each. Dense, specific, mostly rules and examples rather than prose. What you’re building is the briefing document a new colleague would read before their first day working with you. Except this colleague reads it before every single session, without being asked. ### The best way to build your personal context and instructions

The best way I’ve found to build it is to let Claude interview you. Open a fresh Claude Chat session and paste this:

``` shiki “I want you to build my personal context for me. Start by asking me about my profession and what I use Claude for. Then, based on what I tell you, ask 6-8 targeted follow-up questions specific to my role — make sure to include: what Claude gets wrong most often when I use it, what would make me immediately reject an output, what good output looks like in my world, what my non-negotiables are, and who I’m ultimately producing work for. Use the AskUserQuestion feature to ask each question. When you have enough, generate a clean instruction block I can copy into my context files.” ```

I absolutely love the AskUserQuestion feature. It makes the interview process feel natural and conversational. Claude surfaces things you wouldn’t think to write down. When it’s done, save the output as a “.md” file in your Cowork folder. **Your context can live in two places:** 1. **Context** **files** (“.md” files in your folder) - my preference, easier to version control and edit 2. **Instructions** **panel** (top-right of your project) - same mechanic, just a different interface

Same content, same loading behavior, your choice of where to store it. That’s the foundation. *Next: Skills, Connectors, Scheduled Tasks, and Dispatch.*

------------------------------------------------------------------------

## Claude Cowork Skills: Reusable Workflows on a Single Command

## Continue reading this post for free, courtesy of Ilia Karelin.

Claim my free post

Or purchase a paid subscription.

Previous Next