Claude Cowork: the three MCP-backed routines that make the install non-negotiable
Cowork hit general availability this week. These three routines lean entirely on MCP connections to Calendar, Gmail, HubSpot, and Google Drive. Twenty minutes to set up. A week of measurable time back.
Claude Cowork hit general availability this week. Most of the coverage will frame it as "Claude Desktop but it can do stuff for you." That framing will cost you the first month.
The mental model that makes Cowork productive immediately is different. Cowork is the first consumer-grade AI surface where the model does not need you to hand it context. Through MCP connections, it reads your calendar, your inbox, your CRM, and your drive on its own. Your job is to tell it the outcome you want and review the result.
Everything below is built on that premise. If you find yourself pasting data into Cowork, stop. Connect the source instead. That one shift is worth more than any prompt you will learn this year.
What Cowork actually is, technically
Cowork is a local macOS or Windows application that hosts Claude and a set of workspace primitives: a folder it can read and write, a sandboxed shell, a browser it can drive, and a catalog of MCP servers you opt into. The productivity delta over Claude Desktop is not intelligence, it is reach. A Cowork session that has your Google Calendar, Gmail, and HubSpot MCP connections enabled can answer "what is on my plate this week and which deal is slipping" in one turn. The same question in Claude Desktop without connectors requires you to assemble three exports.
This article walks through three routines that lean hard on that reach. Each takes under five minutes to run once the connections are in place. Together they replace roughly an hour of Monday-morning assembly work.
Where the minutes go (self-reported, n=14 executives, week of April 20)
Routine 1: The Monday briefing
This is the one that earns the install by 9:03am Monday. No pasting, no attaching, no exporting. You ask and Cowork reads.
Connections required: Google Calendar (or Outlook), Gmail (or Outlook), HubSpot (or Salesforce/Attio), Google Drive (or OneDrive).
Prompt:
Build my Monday briefing for the week of [date]. Pull directly from my connected tools. Produce a single page with these sections in order:
1. The three meetings this week where I need to decide, not discuss. Pull from Calendar. Flag any where the attendee list has grown by more than two people since last week.
2. The five emails from last week I did not reply to that are still open. Pull from my Inbox, filter for threads with no outbound reply from me, sent within the last seven days, from senders outside my company.
3. Deals that changed stage last week, sorted by revenue impact. Pull from HubSpot. Flag any that moved backwards.
4. The one document in my Drive I said I would read this week but have not opened. Cross-check against my calendar notes.
5. The single highest-leverage thing I could do on Monday before noon.
No preamble. Cite the source for each item. If a connection returned nothing, say so.
What you will see on the first run: Cowork fires a handful of tool calls, announces what it is doing ("Reading calendar for 4/27-5/3... Searching inbox for unanswered threads..."), and produces the page in under ninety seconds. Read it. Edit one line. That page is your week.
Routine 2: The MCP inbox triage
This is where I most commonly see people revert to 2024 habits. The instinct is to copy subject lines into a document. Do not. Cowork reads Gmail or Outlook through the connector; there is nothing to paste.
Connections required: Gmail or Outlook. Optionally HubSpot, for the CRM-match step below.
Prompt:
Triage my last 72 hours of inbox. Read directly from Gmail.
Classify every thread where I am in the To or CC line into one of:
• Reply today (direct ask, customer, or anything with a deadline under 48 hours).
• Reply this week (context required, but not urgent).
• Delegate (name the person on my team to hand it to).
• Archive (FYI, newsletter, or already resolved).
For every item in "Reply today," draft a two-sentence reply in my voice. Do not send. Output as a numbered list with subject line, sender, classification, and draft if applicable. Then check the sender email against my HubSpot contacts and flag any that match an active deal.
Three things Cowork does here that Claude Desktop cannot. It reads the whole thread (not just the latest message), which dramatically improves the triage accuracy. It checks HubSpot cross-reference on its own, which is the difference between "inbox triage" and "deal-aware inbox triage." And it writes drafts to a file in your workspace folder that you can open in a real editor if one needs serious editing.
Critical rule: do not let Cowork send replies. The Gmail → Send permission is off by default and should stay off. Your workflow is "draft in Cowork, paste into Gmail, send with your face on it." Every week someone tries to turn this on and every week they regret it.
Routine 3: The deal pulse check
This one changes Mondays more than the other two combined, because it sees across a data set no human actually reads in full.
Connections required: HubSpot, Salesforce, or Attio. Optionally Granola, Fireflies, or Zoom AI Companion for call-sentiment overlay.
Prompt:
Run a deal pulse check for my book. Pull directly from HubSpot.
For every open deal with close date in the current or next quarter:
• List the three most recent activities (calls, emails logged, notes) with date and summary.
• Flag deals where the last activity is more than 10 days old.
• Flag deals where the most recent call transcript (Granola) shows customer sentiment that dropped from the prior call.
• Rank the top five at-risk deals with one-sentence reasoning.
• Draft a two-sentence internal Slack message to the AE on each of the top three, naming the one action I want them to take this week.
Output a table plus the three Slack drafts underneath.
Executives in a revenue seat should run this every Monday at 7:45am. The model catches things you would not. Specifically: drops in call sentiment are a two-week leading indicator of stall, and nobody reviews them manually. The Granola plus HubSpot combination gives you that signal for free.
Cowork at-risk-deal detection vs manual review (pilot, n=62 open deals)
Guardrails and what not to connect
A few hard-won rules.
- Never turn on Gmail Send. Cowork's draft is almost always good. Your send button is load-bearing. Keep it.
- Never give HubSpot write permissions on the first week. Read-only is enough for every routine above. Earn trust with the model before letting it change deal stage.
- Do not connect your personal inbox and work inbox in the same project. The triage classifier mixes them and you will end up drafting a reply to your mortgage broker in your VP of Sales's voice. I have done it.
- Keep the workspace folder per project, not per year. Every big project deserves its own folder with its own connected set. A global "Cowork" folder collapses into noise in four weeks.
- Revisit connectors monthly. New MCP servers ship weekly. If Linear or Notion or 1Password adds value to a routine, connect it. If a connector is unused for 30 days, turn it off.
Moving from beginner to expert: what changes in month two
The first month of Cowork is about replacing manual assembly with connector pulls. That is the 70% win. The next 30% is procedural. In month two, three things start to matter more than the prompts themselves.
First, named project folders. Give each big workstream its own Cowork project folder with its own connected set and its own "standing instructions" file. The standing instructions tell Claude your role on this project, your vocabulary, and what "good" looks like. Every prompt in that folder inherits that context.
Second, recurring tasks. The three routines above become scheduled tasks, not manual prompts. Cowork's scheduling system fires the Monday briefing at 7:30am without you asking. You wake up to it.
Third, outbound drafts. The endgame of the inbox routine is not triage. It is that every morning you have two or three drafts waiting that you can send with a single keystroke after a quick edit. That is the shift from "I use AI" to "AI is where my work starts." Once that shift is complete, you are not going back.
What to do this week
- Install Cowork from claude.ai if you have not. Budget 10 minutes.
- Connect Google Calendar, Gmail, HubSpot, and Google Drive under Settings → Connectors. Budget 10 minutes.
- Run Routine 1 on Monday morning. Edit the output. Send yourself the final version by email.
- Run Routine 2 before lunch. Draft three replies in Cowork. Paste into Gmail and send with your face on it.
- Friday: run Routine 3. If it flagged a deal your team did not know was at risk, the install has already paid for itself.
Everything else you read about Cowork this month will be about novel use cases. Those come later. The three above are the routines that make the install non-negotiable by Friday, and they work because the model reads your work directly. Every minute they save comes from that one architectural fact.
Pick the next useful thing.
Build a Safe vs Risky AI Chatbot Detector Game with Your Kid
A 60-minute family activity that teaches kids to spot risky chatbot answers with zero screens required for the core lesson.
Turn Apple Watch Sleep Data into One Better Week with GPT-5.5
A five-minute Sunday ritual using Apple Watch sleep data and GPT-5.5 to pick one practical behavior change.
The $65 Billion Anthropic Bet: What It Means for Your Stack
What Google and Amazon investment means for pricing, tooling, and your 2026 agent roadmap.
Three deep dives. Four useful moves. One email worth opening.
PromptHacker turns the AI firehose into practical next steps for work, health, family, and everything time keeps trying to steal.
No comments yet