Claude Cowork: An Onboarding Guide
21 June 2026
Most people open an AI tool and start typing — then wonder why the output sounds generic and the work doesn’t stick. Cowork is different: it works with your local files, runs longer tasks, and behaves closer to an agent than a chat box. Set up a small structure once, and it pays back every session. Here’s the whole thing, in about 30 minutes.
01Create the folder structure
Cowork reads local files — everything stays on your computer, nothing goes to a server.
Claude-Work/
├── Context/
│ ├── about-me.md
│ ├── work-preferences.md
│ └── current-focus.md
├── Projects/
├── Outputs/
├── Templates/
└── CLAUDE.md- Context/ — who you are and how you work. Claude’s memory of you; without it every session starts from zero.
- Projects/ — input materials: transcripts, notes, PDFs, data.
- Outputs/ — where finished work is saved.
- Templates/ — examples of good output, so Claude sees what “quality” means to you.
02Write CLAUDE.md — the behaviour rules
This file loads before every task. It defines what Claude can do, can’t do, and how to work with you.
- Behaviour: plan before starting · ask if unsure · never delete without permission.
- Output: save in
Outputs/· Markdown for documents ·session-notes.mdat the end. - Boundaries: no internet unless asked · no changes outside the agreed folder · if unsure, stop and ask.
Keep it under ~2000 words — every word costs tokens on every task. Short and precise beats long.
03Create the 3 context files
This is what separates generic AI output from text that sounds like you.
The easy way: open Cowork, connect your folder, and write — “Interview me one question at a time and create about-me.md, work-preferences.md and current-focus.md.”
- about-me.md — your role (not a biography), how you define good work, your thinking style.
- work-preferences.md — preferred format, where/how to save files, when to ask vs. act.
- current-focus.md — active projects, this period’s priorities, what you’re not doing.
04Set Global Instructions & choose your model
Settings → Cowork → Edit next to Global Instructions. Configure once; applies everywhere.
Set language & tone, “read about-me.md first”, structured format, and “always ask before
deleting/sending.”
- Sonnet — daily driver. Fast; great for writing, organising, analysis; saves your limit. ~80% of tasks.
- Opus — complex tasks: deepest reasoning, strategy, bigger documents. Slower, uses more limit.
05Add the plugins for your work
Each plugin brings ready-made skills for a domain. Install once, use every session: Sales (CRM, proposals) · Legal (contracts, compliance) · Analytics (data, reports) · Communication (email, Slack, meetings) · File Manager (organise, batch rename) · Research (search, synthesis).
Chat vs. Cowork
Claude Chat — plan & write
Planning before action · quick questions & brainstorming · writing/copywriting that needs no files · preparing a clear prompt to hand to Cowork.Claude Cowork — execute & files
Working with local files · long, complex tasks · creating real Excel/Word/PPT · automated tasks via/schedule while you do other things.The 7 behaviour rules
- Start with the outcome, not the task. Not “make a report” but “a one-page PDF competitor analysis to pitch an investor Wednesday.”
- Use Projects for multi-session work. More memory; pick up where you left off.
- Ask for a plan before execution. Easier to fix a plan than a finished 20-page document.
- Write session-notes at the end. Next time start with “Read session-notes.md.” Zero lost context.
- Add only relevant files. Every file loaded costs tokens.
- Don’t bring chat habits into Cowork. Chat is for questions; Cowork is for execution.
- Plan in Chat, build in Cowork. Exploring inside Cowork is expensive.
● Living guide — updates
- 2026-06-21 — First published.
This guide is updated as Claude Cowork evolves.