Most people use Claude like a conversation. You type a question, it answers, you ask another one. Thats chat. Cowork is different. Cowork is you handing a task to someone, walking away, and coming back to finished work in your folder. Thats delegation. Completely different relationship.
The moment that distinction clicks you stop asking it questions and start giving it work.
Cowork reads and writes files directly on your hard drive. Do not give it access to everything. Make a folder on your desktop called "sandbox." Thats where it lives. Everything outside that folder is untouchable. Grant permissions only for that folder.
The model matters too. Sonnet 4.6 for 99% of tasks — its cheaper and handles almost everything. Opus 4.6 only for complex high-stakes reasoning. I call this the Einstein rule — dont put Albert Einstein on kitchen duty. Keep extended thinking on so it actually processes logic instead of pattern-matching.
First test: drop a dozen mixed invoices into the sandbox. "Sort these by category and generate an Excel summary." It outlines a plan and executes. Your invoices are organized before you finish your coffee.
Without context, Cowork has no idea who you are. You explain yourself from scratch every session. Waste of time.
The fix: .md files sitting in your project folder that Claude reads before every prompt.
about_me.md — who you are, what your business does, who your customer is, how you make money, current priorities. Claude reads this every time.
brand_voice.md — how you communicate. Tones you use. Phrases you hate. Examples of your actual writing. This stops you from sounding like everyone elses Claude output.
working_preferences.md — how you want tasks managed, where files get saved, output formats.
Dont write these from scratch. Ask Claude to interview you. "Ask me questions and use my answers to build my business brain files." 15 minutes. Saves hours every week.
Each layer builds on the last. Dont try to build the full machine day one. Automate one weekly task first. When thats solid, chain two together into a plugin. Then schedule it. Then dispatch it from your phone while youre somewhere else.
If your app isnt natively supported by Claude, Zapier's MCP connects to over 8,000 applications. HubSpot, Airtable, whatever youre running. Build an MCP server inside Zapier, select the tools and actions you want, paste the URL into Claude's connector library. Done. 10 minutes.
This single integration kills the entire "but my tool isnt supported" objection. Permanently.
The Apify MCP is the other one nobody talks about. Want to scrape data from YouTube, TikTok, or Instagram? Create an Apify API token, connect it in Cowork. Claude sorts through 1,300+ scrapers and deploys the right one automatically. No manual configuration.
Morning brief. Schedule Claude to build your daily dashboard before you sit down. Calendar summary, outstanding emails, local weather, industry news. It drafts reply emails for you to review. You sit down and everything is already done.
Content repurposer. Give Claude a YouTube URL. It extracts the transcript, adds it to a Notion page, writes platform-specific posts for LinkedIn and X. One input, three outputs, zero manual work.
Financial dashboard. Monthly scheduled task. Feed it your transactions or receipts. It categorizes expenses, checks balances, generates an HTML dashboard showing profit and loss. Your accountant gets a clean report. You spent zero minutes on it.
Every word costs tokens. Three mistakes that drain your limits fast:
Context rot. Talking in the same chat for hours. Claude reloads the entire conversation history every message. Finish one topic, open a new window. 30-45 minutes per session max.
Too many connectors. Every active MCP eats tokens before you type anything. Only turn on what the current task needs.
Processing in conversation. If you need to process 100 invoices, dont have Claude read each one in chat. Have it write a script (a skill) that processes them. The script uses a fraction of the tokens.