✦ April 2026 Update

Claude Cowork
Setup Guide

The complete 2026 guide to Claude Cowork by Ruben Hassid — folder structure, core identity files, Global Instructions, Wispr Flow voice dictation, and token-saving strategies.

By Ruben Hassid · How to AI · 448,000 subscribers · 20-min setup
What is it?

Claude Cowork — Best AI Feature for Non-Developers

🔗

Anthropic's Claude Cowork is a desktop app mode where Claude works directly inside a folder on your computer — reading context files, executing tasks, and saving deliverables automatically. As of April 2026, Anthropic adds $323.5M in ARR per day, surpassing ChatGPT in revenue.

🗂️
Folder-Based Workspace
Cowork operates inside a selected folder on your computer. It reads ABOUT ME files before every task and saves all outputs to OUTPUTS.
🎙️
Voice-Driven Workflow
Paired with Wispr Flow, you speak at 150 wpm instead of typing at 60 — producing richer, more natural context for better outputs.
💡
Persistent Identity
Three small files in your ABOUT ME folder tell Claude exactly who you are every session — no re-explaining required.
I — Folder Architecture

Your Cowork Folder Structure

🔗

The entire power of Claude Cowork lives in how you structure your folder. Three subfolders, three core files, total under 6,000 tokens.

📁 Claude Cowork/
📁 ABOUT ME/ ← read before every task
📄 about-me.md — who you are (under 2,000 tokens)
📄 anti-ai-writing-style.md — 80+ banned AI words
📄 my-company.md — goals & strategy (under 1,000 tokens)
📁 OUTPUTS/ ← Claude saves deliverables here
project-name/ ← one subfolder per project
📁 TEMPLATES/ ← reusable output skeletons
saved automatically when you say 'Save as template'
II — The Three Core Files

What Goes in ABOUT ME

🔗

These three files are the only ones Cowork reads automatically. Keep them lean — every extra token costs you on every single session.

🧑
Who you are, how you think, how you want Claude to write for you. Generated via 20-question interview with AskUserQuestion. Target: under 2,000 tokens. Trim raw transcripts — keep only extracted patterns.
✂️
Your taste as a ruleset. Bans 80+ AI words (delve, harness, tapestry), kills reframe patterns, limits paragraphs to 3 sentences. Without it, Claude writes like Claude. With it, Claude writes like you.
🎯
Your north star. Targets per platform, services/offerings, and 'what I'm saying no to.' Under 1,000 tokens. Update quarterly or when priorities actually change — not on a schedule.
III — The Secret Sauce

Global Instructions

🔗

Found at Settings → Cowork → Edit Global Instructions, this persistent prompt tells Claude which files to read before every task, where to save deliverables, and what to do with unclear briefs — without you ever repeating it.

📋
What to include
Tell Claude to read all files in ABOUT ME/ before every task. Instruct it never to read OUTPUTS/ or TEMPLATES/ unless explicitly pointed there. Always save deliverables to OUTPUTS/ under a project subfolder.
Why it matters
If your context window stays lean (files under 6,000 total tokens), Claude reads them fully and precisely every session. Exceed that, and Claude summarises loosely — losing the precision that makes Global Instructions valuable.
IV — The Bottleneck is You

Wispr Flow Voice Dictation

🔗

Cowork reads 100,000 words in 15 seconds. It waits for you to type at 60 wpm. Wispr Flow lets you speak at 150 wpm with near-perfect accuracy — anywhere on your computer, including inside the Cowork chat box.

💬
Initial Prompts
Speak your task instead of typing. Speaking gives 3–4× more context naturally — you're a yapper by nature, and more context produces better outputs from Claude.
📝
When Cowork generates its interactive form, click the options — then dictate the custom 'Other' answers that contain the nuance Claude really needs.
🔄
Feedback & Pivots
When output is off, speak rich feedback: 'The tone is too stiff. Sound like I'm texting a friend who runs a 200-person company. Keep the data, redo only section 2.' Spoken feedback is richer than typed.
V — Model Selection

Choosing the Right Claude Model

🔗

Smarter = more expensive. Match the model to the task complexity to preserve your token budget for work that truly needs it.

🧠 Opus 4.6
Complex tasks, strategic thinking, long-form drafts. Use with Extended Thinking for the highest quality. Most expensive.
⚡ Sonnet
Grammar checks, brainstorming, formatting, short answers, quick drafts. Fraction of Opus cost. Good daily workhorse.
🌸 Haiku
Simplest, cheapest tasks. Routing, classification, single-step processing. Reserve for truly routine operations.
VI — Token Optimisation

6 Ways to Save Credits in Cowork

🔗

Every message causes Claude to re-read the entire conversation. At 30 messages, that's 232,000 tokens burned on history alone.

🔁
Restart conversations early. When something goes wrong, click 'Restart the conversation from here' on an earlier message — not at the bottom. Don't type follow-ups; they stack all history.
🔢
Fresh session every 20 messages. One developer found 98.5% of tokens went to re-reading history; only 1.5% to actual output. Summarise, copy, new session, paste summary.
📦
Batch your tasks. Three separate prompts = three full context reloads. One prompt with three tasks = one reload. Always combine related instructions into a single message.
Use Sonnet / Haiku for simple tasks. Save Opus 4.6 for complex, high-stakes work. Using cheaper models on routine tasks frees up 30–70% of your budget.
🗜️
Keep ABOUT ME files small. Cowork reads them before every task. Ruben's about-me.md went from 22,000 tokens to under 2,000 — same signal, 10× less noise.
🕐
Spread usage across the day. Claude uses a rolling 5-hour window. Split sessions morning/afternoon/evening so earlier usage rolls off. Avoid peak hours 5–11 AM Pacific on weekdays.
VII — HowTo

Set Up Claude Cowork in 20 Minutes

🔗

Block 20 minutes this week. Bonus: do it as a team. Here's the exact sequence:

1
Go to claude.com/download, download the desktop app, and subscribe to the Pro plan ($20/month) or Max ($100/month). Open the app, click the Cowork tab, and select your folder.
2
Make a folder named 'Claude Cowork' on your computer. Inside: create ABOUT ME/, OUTPUTS/, and TEMPLATES/ as empty subfolders.
3
In a Cowork session with Opus 4.6 + Extended Thinking, use the interview prompts from the article to generate about-me.md, anti-ai-writing-style.md, and my-company.md.
4
Settings → Cowork → Edit Global Instructions. Delete existing text. Paste the provided template telling Claude to read ABOUT ME before every task and save to OUTPUTS.
5
Go to wispr.ai, download Wispr Flow, install it, and pick a keystroke (e.g. Shift). Hold key → speak → release. Your words appear anywhere on screen. Free tier: 2,000 words/week.
6
Open Cowork, speak your first real task using Wispr Flow. Answer AskUserQuestion forms by speaking. Review the output — you'll feel the difference in the first 3 minutes.
7
After a successful session, say: 'Save this as a template in TEMPLATES/.' Claude strips content, keeps structure. Reference it by name in any future session to reuse the format.
VIII — FAQ

Frequently Asked Questions

🔗
What is Claude Cowork?

Claude Cowork is a desktop app mode by Anthropic where Claude works inside a selected folder on your computer — reading your personal context files, executing tasks using tools, and saving deliverables automatically. Designed for non-developers.

What plan is required?

At minimum the Claude Pro plan ($20/month). For heavy usage without hitting limits, the Max plan ($100/month) is recommended.

Why must ABOUT ME files stay under 6,000 tokens?

Claude reads these files before every session. Exceed 6,000 tokens and Claude starts summarising them loosely instead of reading precisely — losing the accuracy that makes Global Instructions powerful. The context window should serve your actual task, not your profile.

How does the TEMPLATES folder work?

After a successful output, say 'Save this as a template in TEMPLATES/.' Cowork strips the content but preserves the skeleton — sections, order, format, length. Reference it by name in future sessions.

Why is restarting the conversation the biggest token hack?

Every message causes Claude to re-read the entire conversation history. At ~500 tokens per exchange: 20 messages = 105,000 tokens; 30 messages = 232,000 tokens — mostly on re-reading, not output. Restarting from an earlier point cuts this exponential cost.

What is Global Instructions?

A persistent prompt at Settings → Cowork → Edit Global Instructions. It runs before every task, telling Claude what to read, where to save, and how to handle unclear briefs — without you repeating yourself each session.

Is Wispr Flow free?

Yes — Wispr Flow's free tier allows 2,000 words per week. It works anywhere on your computer by holding a chosen keystroke. Near-perfect accuracy, no app integration needed — it types wherever your cursor is.

What model should I use for complex tasks?

Use Claude Opus 4.6 with Extended Thinking for complex, high-stakes tasks. Use Sonnet for drafts and formatting. Use Haiku for the simplest tasks. Smarter = more expensive.

How do I create about-me.md from scratch?

Open a Cowork session with Opus 4.6 + Extended Thinking, paste the interview prompt from the article, answer 20 AskUserQuestion prompts by dictating with Wispr Flow. Cowork compiles into about-me.md under 2,000 tokens.

What is the rolling 5-hour window?

Claude's token usage resets on a rolling 5-hour basis. If you burn your limit in one morning session, your capacity sits unused for the rest of the day. Splitting into 2–3 daily sessions lets earlier usage roll off and effectively expands your daily budget.

What is task batching?

Instead of sending three separate messages ('Summarise this.' then 'List main points.' then 'Suggest a headline.'), combine them into one: 'Summarise this, list the main points, and suggest a headline.' Three separate messages = three full context reloads. One batched message = one reload.

Does Cowork read OUTPUTS automatically?

No. OUTPUTS is only read when you explicitly ask: 'Read the report in OUTPUTS/project-name.' This keeps your context window clean and your token budget intact.

IX — Glossary

Key Terms

🔗
Units of credit consumed by Claude. Every message re-reads full history; long conversations burn tokens exponentially.
The maximum text Claude can hold in mind at once. ABOUT ME files, history, and task content all compete for this finite space.
Built-in Cowork tool that generates interactive forms with multiple-choice options and free-text 'Other' fields for gathering task context.
Persistent Settings prompt that runs before every Cowork task. Tells Claude which files to read and where to save deliverables.
Optional Claude reasoning mode for deeper, more considered outputs. Best paired with Opus 4.6 for complex tasks. More tokens, more time.
Saved output skeleton in TEMPLATES/. Claude strips content, preserves structure. Referenced by name in future sessions.
Claude's usage limit resets on a rolling 5-hour basis. Spread sessions to let earlier usage roll off and expand daily capacity.
File banning AI clichés (delve, harness, tapestry), reframe patterns, and formatting rules. Makes Claude write like you, not like AI.
Feature to restart from any earlier message, eliminating exponential token costs of long conversation history. The single biggest token-saving hack.
Combining multiple instructions into a single message. Three separate prompts cause three context reloads; one batch causes one.