STEP 4 · REAL WORK

Claude in Action

Copy-paste prompts and complete setups from people who use Cowork every day. No theory, just what works.

Before You Start These workflows assume you've completed the Claude Beginner setup, or at minimum: Claude Desktop installed, a workspace folder selected, and Gmail + Calendar connected. If you haven't done that yet, start there and come back.

Eight Workflows

1
Workflow 1

The 6 AM Dispatch

Queue 5 tasks from your phone before breakfast. Review finished work over coffee.

Who This Is For

CEO or founder managing multiple roles. You have 90 seconds before your day explodes. You need to queue work from anywhere, not just at your desk.

The Setup

Organized Workspace Folders

Create: Active Work → Today's Work. This is where Claude queues tasks and outputs pile up. You review from this single folder.

Gmail + Calendar Connectors

Connect your email and calendar. Claude can now read incoming messages and your day's schedule to prioritize intelligently.

Claude on Your Phone

Install Claude on your phone. As of the July 2026 mobile release, you can start the whole session from your phone at 6 AM and it runs in the cloud — your Mac can be closed. The catch: a cloud session can't touch your local desktop folders. Route these outputs to a connector (Drive, Notion) if you want to review them on your phone. See the note below.

The Prompts

Go through my emails, prioritize the tasks for me that are most important, and organize them. Draft emails where you can to the ones that I should be responding to.
Plan out content for Substack. Make a branded HTML with candidate ideas. This month I want to speak about [TOPIC]. Search the web, look for compelling stories I may want to cover, and organize that as a content calendar.
Organize my Downloads folder. Segment [PROJECT] relevant files into that folder. Within the [PROJECT] folder, ensure you have relevant files categorized by customer and task.

What Claude Produces

Prioritized email drafts ready to send. A content calendar in HTML with story ideas, sourced and themed. Files organized by project and task type, all before you sit down at your desk.

Why It Works You fire 5 tasks from your phone in 5 minutes and they run in the cloud — no Mac left running, no laptop open. You queue, walk away, review outputs over coffee from any device. No waiting. No context switching.
One-Way Mirror: Where This Work Lives A session you start on your phone runs in the cloud and shows up everywhere — phone, web, and desktop. But work you do in a local desktop session stays on the desktop; your phone never sees it. Rule of thumb: the cloud is visible everywhere, your desktop is a one-way mirror. So if a task writes to a local folder (like "Today's Work"), you'll only see those files on your Mac. Point the output at a connector — Google Drive, Notion — and the finished work follows you to your phone.
Make It Yours Replace email triage with your biggest morning time sink. Swap 'Substack' for your platform. Change 'Downloads' to whatever folder collects clutter. The pattern works for any role that has repeatable morning work.
Level Up: Turn This Into a Skill If you're running this dispatch every morning, stop copy-pasting the prompt. Tell Claude: "Build me a skill called 'morning-dispatch' that runs my 5 morning tasks. Read my current dispatch prompt and convert it into a reusable skill with my email rules, content platform, and folder structure baked in." Once it's a skill, one sentence triggers the whole thing, and it gets smarter as your MEMORY.md grows. New to skills? See Claude Advanced, Section 1. Scheduled tasks? Section 6.

Inspired by Christian Pean MD, Techy Surgeon

2
Workflow 2

The Essential Three, Scheduled Tasks Starter

Three automated tasks that cover 80% of your daily overhead.

Who This Is For

Anyone starting with automation. You don't need ten tasks. Start with three. Let them run for a week and feel the difference before you add anything else.

The Setup

Gmail + Calendar Connectors

Connect your email and calendar so Claude can read them as part of each scheduled task.

Scheduled Tasks Enabled

Make sure Cowork has permission to run scheduled tasks. This happens during setup, but verify in Settings if tasks don't fire.

The Three Tasks

1

Morning Briefing (Daily 8 AM)

Summarize my unread emails by priority. Check today's calendar. Create a prioritized task list with talking points for each meeting. Output as an HTML file in my workspace.
2

Meeting Prep (Daily, 30 min before first meeting)

Review the agenda and attendees for my next meeting. Pull any recent email threads with these people. Draft 3 talking points and 2 questions to ask.
3

Evening Debrief (Daily 6 PM)

Summarize what I accomplished today based on sent emails and calendar events. Flag anything that needs follow-up tomorrow. Keep it under one page.

What Claude Produces

A morning HTML briefing with your day's priorities. Pre-meeting prep with talking points and questions. End-of-day summary with follow-ups flagged and ready for tomorrow.

Why It Works Start small. Three tasks. One week. You'll feel the difference before you add anything else. These three cover: morning focus, meeting readiness, and end-of-day clarity. That's the foundation.
These Run in the Cloud — Laptop Can Be Closed Since the July 2026 release, scheduled tasks run in the cloud on Anthropic's servers. Your Mac can be asleep or shut. The three tasks here are pure-connector (they read Gmail and Calendar and write an HTML briefing), so they fire on schedule and land on your phone whether or not you opened your laptop. The one exception: a task that reads or writes a local desktop folder still needs the Mac awake and the desktop app open, because the cloud can't see your local files.
Add a Freshness Check Put one line at the top of every scheduled task: before you run, confirm the input still exists and has new data since last time. If nothing changed, say so and stop. This kills the quiet failure where a task runs for weeks over a dead folder and nobody notices. More on this in When Automations Go Wrong, below.
Make It Yours Swap meeting prep for your recurring prep type. Change debrief time to your cadence. Cron examples: 0 9 * * 1-5 for weekdays at 9 AM, 30 8 * * 1 for Mondays at 8:30 AM.
Level Up: Wrap Your Tasks in a Skill You've got the scheduled tasks running, now make them smarter. Tell Claude: "Look at my 3 scheduled tasks. Build a skill called 'daily-ops' that combines the briefing, meeting prep, and debrief into one system. Include my email priority rules, meeting types, and output preferences so I don't have to re-specify them." A skill gives your scheduled tasks a shared brain, consistent formatting, your preferences baked in, and better outputs over time.

Inspired by Christian Pean MD, Techy Surgeon

3
Workflow 3

Client Deliverable Factory

Turn raw briefs into polished reports every time.

Who This Is For

Consultant, lawyer, agency professional. You produce the same type of deliverable repeatedly. You want consistent quality and fast turnaround.

The Setup

Project Folder with Template

Add your report template (.docx). Include 2 strong past examples as reference. Add the client brief and any reference docs. Include project instructions: "Include executive summary. Never exceed 15 pages. Structure: summary → findings → recommendations → next steps → appendix. Output as .docx."

The Prompt

The client just sent their brief — it's in the project folder. Read the brief, compare it to our template and past examples, then ask me 3-4 clarifying questions before you generate the full report as a Word doc.

What Claude Produces

A templated Word document with executive summary, findings, recommendations, next steps, and appendix. All formatted to your style, based on the template and examples you provided.

Why It Works Template + examples stay permanently loaded in the project. Scoped memory means Claude improves with each report. The third report is noticeably better than the first.
Make It Yours Swap template for proposals, SOWs, legal memos, case studies. Add letterhead .docx for auto branding. Key: give Claude 2-3 strong examples of what 'good' looks like for your domain.
Level Up: Build a Deliverable Skill If you produce the same type of document more than twice, it should be a skill. Tell Claude: "Build me a skill called 'client-report' that generates [DOCUMENT TYPE] from a brief. Read my template and the last 3 reports I produced. Bake in the structure, formatting rules, and quality checks. The skill should ask me clarifying questions before generating, just like a good junior analyst would." Now one sentence, "run client-report for Acme", produces a first draft that matches your standards.

Inspired by Ruben Hassid, How to AI

4
Workflow 4

Deep Research Engine

Verified, citation-rich intelligence on any person, company, or market.

Who This Is For

Founder prepping for investor meetings. Analyst sizing a market. Sales rep researching accounts. Anyone who needs research they can actually trust.

What Makes This Different

Most people ask Claude to 'research X' and get a generic summary with no sources. This approach uses a recursive research loop, the same methodology behind dedicated research skills that produce 15-30 page cited dossiers. The difference is structure: Claude decomposes the research into sub-questions, probes sources in priority order, tracks confidence on every claim, and separates evidence from inference.

The Research Loop

Step What Happens
Probe Search sources in priority order: official site → social channels → reviews → press/third-party
Decompose Break each section into 3-6 sub-questions. No claims without citations.
Recurse For each sub-question, inspect sources, write a mini-brief with citations + confidence
Synthesize Combine mini-briefs into the final section. Resolve conflicts between sources.

Evidence Rules

1

Every Claim Cited

Inline citations with source, date accessed, and confidence (High / Medium / Unconfirmed). No naked claims.

2

Three-Tier Separation

Company-stated claims vs. third-party corroboration vs. inference. Each labeled explicitly.

3

No Speculation

If not publicly disclosed, write "No public disclosure found as of [date]" instead of guessing.

The Prompts

Person Research:

Research [PERSON NAME] at [COMPANY]. I'm meeting them for [PURPOSE]. Probe their LinkedIn, company site, talks, podcasts, and social media. Decompose into: career timeline, current role and priorities, communication style, personal interests, and fit with my ask. Cite every claim with source and confidence level. Separate evidence from inference. Output as a Word doc with an executive brief I can read in 5 minutes before the meeting.

Company Research:

Research [COMPANY] — [WEBSITE]. I'm evaluating them as [CONTEXT: competitor / partner / investment]. Probe their site first (product, pricing, about, blog, terms), then social, reviews, and press. Decompose into: business model, product offering, target customers, GTM motion, partner ecosystem, competitive landscape (8-12 players), and momentum in the last 24 months. Cite everything. No speculation on revenue or internal metrics. Output as a Word doc with executive summary.

Industry/Market Research:

Research the [INDUSTRY] market. Probe industry associations, research firms, company filings, government sources, and trade press. Decompose into: market sizing (TAM/SAM/SOM), value chain, competitive landscape, customer segments, technology trends, regulation, risks, and strategic opportunities. Present conflicting estimates as ranges. Cite every figure. Output as a Word doc with a 'So What' section at the end.

What Claude Produces

A 10-25 page Word document with inline citations, confidence scores, a source pack (12-25 annotated links), comparison tables, and an executive summary. The output quality depends on the prompt structure, the recursive loop and evidence rules are what separate this from a generic 'research X' request.

Why It Works The recursive loop forces depth. Decomposing into sub-questions prevents Claude from writing generic summaries. Source priority order ensures you're building from primary sources, not regurgitating blog posts. Confidence scoring tells you what to trust and what to verify yourself.
Make It Yours These prompts are starting points, customize the decomposition list for your domain. Add 'focus on [SPECIFIC ANGLE]' to narrow the research. Drop existing research into the project folder for Claude to build on. For recurring research (weekly competitive intel, regulatory monitoring), convert to a scheduled task. Install the person-research, company-research, or industry-research skills for the full automated version.
Level Up: Build Your Own Research Skill The prompts above are the manual version. The real power is building a research skill that encodes YOUR methodology permanently. Tell Claude: "Build me a research skill using this framework: probe sources in priority order (official site → social → reviews → press), decompose each topic into 3-6 sub-questions, recurse through each sub-question with citations and confidence scoring (High / Medium / Unconfirmed), then synthesize into a final report. Separate evidence from inference. Include a source pack with 15-25 annotated links. Output as a .docx with an executive brief on page one." Once built, saying "research [person/company/market]" triggers a 15-30 page cited dossier, no prompt engineering required. For recurring intel (weekly competitor scans, regulatory monitoring), pair the skill with a scheduled task.

Research methodology by Steve Gustafson

5
Workflow 5

CSV to Dashboard

Turn any spreadsheet into an interactive dashboard without code.

Who This Is For

Anyone with data who needs visuals fast. You have the data. You don't care about the technical implementation. You want a dashboard you can share.

The Setup

Drop CSV/Excel into Project Folder

That's it. No plugins, no connectors, no special configuration. Just add your data file to the project folder.

The Prompt

Read the data file in my project folder. Create an interactive HTML dashboard with: a summary metrics bar at the top, dropdown filters for [COLUMN1] and [COLUMN2], a trend chart showing [METRIC] over time, and a sortable data table below. Use a dark theme. Make it a single HTML file I can share with anyone.

What Claude Produces

A single-file HTML dashboard with Chart.js, dropdown filters, summary cards, responsive design. Fully interactive. Share via email or Slack. Works in any browser.

Why It Works One file, no dependencies, opens in any browser, shareable via email or Slack. Claude handles the Chart.js code, responsive layout, and filter logic. You get dashboards in minutes instead of hours.
Make It Yours Change column names and metrics to match your data. Add 'include a comparison to last month/quarter' for period-over-period. Request chart types: bar for categories, line for trends, scatter for correlations. Combine with a scheduled task for recurring dashboards that update weekly.
Level Up: Make This a Scheduled Task If your data updates regularly (weekly exports, monthly reports), stop doing this manually. Tell Claude: "Create a scheduled task that runs every Monday at 9 AM. It should read the latest CSV in my [PROJECT] folder, generate an updated dashboard with the same layout and filters as last time, and save it as an HTML file with this week's date. Compare to last week's numbers and flag anything that changed more than 10%." Now your dashboard rebuilds itself, and you get anomaly alerts for free.

Which Tool, When

Claude can touch a spreadsheet or a deck several different ways. Reaching for the wrong one wastes time. Use this map, and see the full Claude in Office guide for the why behind each.

You want to…UseWhy this one
Edit a live workbook / model that already existsClaude in Excel (add-in)Only this edits real cells and traces dependencies in your open file.
Generate a fresh, formatted workbook from data or a briefCowork xlsx skillBuilds a new .xlsx from scratch with formulas and styling, no add-in needed.
Edit / restyle an existing deck slide-by-slideClaude in PowerPoint (add-in)Works inside your template; keeps layouts and master styles intact.
Build a new deck from data, with a narrativeCowork pptx / exec-deck skillGenerates a full .pptx from a spreadsheet; the skill bakes in the blueprint→polish method.
A shareable, interactive view of a datasetCSV → HTML dashboardOne self-contained HTML file with filters and charts, opens in any browser.
A designed, on-brand deck or pageClaude DesignImage-led visual polish and working code, when look-and-feel is the point.
A finished invoice / report / receipt to sendPDF agent / invoice-pdf skillLayout, totals, and structure handled; returns a shareable PDF.
Rule of ThumbEditing something that exists → the Office add-in. Generating something new → a Cowork skill. Sharing a view → a dashboard. Making it look designed → Claude Design.
6
Workflow 6

The AI Chief of Staff

The full system, from three tasks to a personalized operating layer.

Who This Is For

Executive managing multiple domains. Anyone who wants the complete automated intelligence system. Start with three tasks. By week four, you have a personal operating system.

The Pattern

This isn't a single workflow, it's what happens when you combine the previous five into a system. Start with the Essential Three (Workflow 2). Add one domain-specific task per week. By week four, you have a personalized operating layer that handles your morning briefing, meeting prep, recurring research, deliverable production, and end-of-day wrap-up, without you re-explaining anything.

The Graduated Rollout

1

Week 1: Deploy the Essential Three

Morning briefing, meeting prep, evening debrief. Just these. Get comfortable with the rhythm. Notice what changes in your day.

2

Week 2: Add One Domain Task

Pick from the menu below based on your role. One task only. Don't add more until you see the value from this one.

3

Week 3: Add a Second Domain Task

By now you see which automations actually save time vs. create noise. Add the second one based on what worked.

4

Week 4+: Customize Based on What's Useful

You now have a personalized operating system. Iterate weekly. Remove what doesn't work. Add what does.

Domain Task Menu

Deal Pipeline Pulse

Scan emails for stale sales conversations. Flag deals going cold. Alert on inactivity over 2 weeks.

Weekly Competitive Intel

Monitor 3-5 competitor sites for pricing, features, blog, and hiring changes. Deliver as a weekly summary.

Customer Success Threads

Compile customer emails and tickets by priority. Flag escalations. Summarize by issue type.

Regulatory Monitoring

Scan relevant government sources weekly for policy changes. Alert on changes that affect your business.

Invoice Cleanup

Scan emails for invoices. Extract amounts, vendors, dates. Flag items over your threshold for review.

The Setup Prompt

Interview me about my professional life. Ask me about my roles, priorities, VIP contacts, deal stages, recurring meetings, and communication preferences. Then generate a comprehensive context file and suggest 5 scheduled tasks based on what you learn.

What Claude Produces

A complete context file + recommended scheduled task library tailored to your actual work. Over time, scoped memory learns your patterns and the outputs get sharper. By week 8, Claude knows your priorities better than most people around you.

"The question is not 'how do I do more?' but 'how do I offload the mechanical work so my limited attention goes to decisions that require judgment?'"
, Christian Pean MD, Techy Surgeon
Why It Works The graduated approach prevents overwhelm. Starting with 3 tasks instead of 10 means you actually stick with it. Adding weekly lets you evaluate what's useful vs. what creates "slop at scale." The interview prompt builds a context file that makes every future task better.
The 'Slop at Scale' Warning Every output requires human review. The value is in first-draft quality and organization, not blind automation. If you're spending more time editing outputs than you saved, that task isn't working. Remove it and try a different one.
Make Your Tasks Learn A scheduled task with no memory repeats itself. It re-reports the same "news" every day because it does not know what it already sent. Fix it with a small state file: the task reads what it reported last time, skips repeats, and logs anything that broke so the next run is sharper. Good automations improve with each run, they do not just run. See When Automations Go Wrong, below.
Make It Yours The domain tasks above are examples, replace them with whatever recurring work eats your time. The interview prompt is the key unlock: it builds a context file tuned to YOUR work, not generic templates. Run the interview again every quarter as your priorities shift.
Level Up: Generate Your Entire Skill Library The interview prompt above is your launchpad. After Claude builds your context file, take it one step further: "Based on the context file you just built, generate 3-5 custom skills for my most repetitive workflows. Each skill should include my preferences, output format, and quality checks. Then suggest which ones should also be scheduled tasks with recommended timing." This turns a one-time interview into a permanent operating system. Each skill runs on one sentence. Each scheduled task runs on autopilot. The context file ties them all together.

Inspired by Christian Pean MD, Techy Surgeon

7
Workflow 7

Spreadsheet → Board Deck

Turn a messy Excel into a board-ready deck, without exporting a single chart.

Who This Is For

Founder building the board update. Consultant turning a model into a client readout. Anyone who has the numbers and needs a deck that tells the story, not a wall of pasted charts.

The Setup

Source Data File

Drop the source .xlsx or .csv into the project folder, or open it directly with the Claude PowerPoint add-in.

Brand Template + Reference Deck

Add a brand template (.pptx) and one strong past deck as a reference so Claude matches your house style.

Optional Project Note

"Board audience. Max 12 slides. Every slide scannable in five seconds." A one-line brief sharpens every step that follows.

The Prompts, Run in Order

Read the data file in my project folder. Act as a presentation consultant. Create a blueprint: objective, audience, the single key message, and the recommended number of slides. Don't build anything yet.
Create a slide-by-slide structure — titles plus the purpose of each slide. Sequence it as a narrative: context → key insights → why they matter → recommended actions.
Generate the deck as a .pptx using my brand template. Pull the real numbers from the data file. One chart per slide max, each tied to the slide's point.
Final pass: reduce text, sharpen every headline into a claim, and make each slide scannable in five seconds.

What Claude Produces

A branded .pptx where the deck has an argument, the charts are sourced from your actual data, and the headlines state conclusions instead of labeling axes. A first draft you'd edit, not rebuild.

Why It Works The blueprint forces a key message before a single slide exists, which is exactly the step that separates an analyst's deck from an "AI deck." Pulling numbers straight from the file kills copy-paste errors. The five-second pass does the tightening you'd otherwise do at midnight.
Make It Yours Swap "board" for investor, client, or all-hands and the structure adapts. Point it at a monthly export and pair with a scheduled task for a deck that rebuilds itself each cycle.
Level Up: Build the exec-deck Skill If you do this monthly, encode it. Tell Claude: "Build me a skill called 'exec-deck' that turns a spreadsheet into a board-ready deck using the blueprint → narrative → in-place generate → story → executive-polish method. Bake in my template, my slide structure, and the five-second-scan check. Ask me objective and audience first." Then "run exec-deck on Q2 numbers" is the whole prompt. New to skills? See Claude Advanced, Section 1.

Method adapted from the Claude × Office playbook (Guillermo Flor, PMF) • structured by Steve Gustafson

8
Workflow 8

The Meeting Follow-Up Engine

The recap email, in your voice, drafted before you close the meeting tab.

Who This Is For

Anyone who runs back-to-back calls and loses the thread on follow-ups. Consultants who owe every prospect a recap. Founders who decide five things in a meeting and remember two by evening. The email that should go out in ten minutes instead goes out in three days, or never.

The Setup

Your Voice Profile (do this once)

Claude writes a profile of how you actually email, then follows it every time. You build it once and reuse it forever. The prompt is below.

Granola, the meeting memory

Download it at granola.ai. It's a notepad, not a bot, nothing joins your call. It transcribes in the background on Zoom, Meet, and Teams, then hands you clean notes when the meeting ends. In Claude, go to Settings → Connectors, search for Granola, and connect it. Now Claude can read any meeting you've had.

Gmail Connector

Settings → Connectors → Gmail. This is how Claude reads your sent mail to learn your voice, and drops the finished draft where you'll find it.

The Prompts, Run in Order

First, build your voice profile (once):

Read my last 30 sent emails in Gmail. Write my voice profile: how I greet, how I sign off, my sentence length, my go-to phrases, the words I never use. Format it as writing instructions to follow when drafting as me.

Save that output as project instructions. This is the step that makes your emails sound like you and not like a chatbot.

Then, after every meeting:

Pull my last meeting from Granola. Draft the follow-up using my voice profile: what we decided, action items with owners, one clear next step. Keep it as short as I keep mine. Create it as a Gmail draft to the attendees.

What Claude Produces

A draft sitting in Gmail, attendees already in the To field, written in your voice: what you decided, who owns what, the single next step. You read it once and send. The email is waiting before your coffee gets cold.

You Still Hit Send Claude drafts. It cannot send. Nothing leaves your outbox without your click. Your clients still talk to you, just faster, and nobody on the other end can tell an "AI email" from a fast one, because there's no such thing. There are emails that sound like you and emails that don't.
Why It Works The voice profile removes the tell. Granola removes the "what did we actually decide" scramble by remembering every meeting for you. And moving the drafting to right-after-the-call, while it's fresh, is what turns a recap you keep meaning to write into one that's already done.
Make It Yours The real unlock isn't one email, it's the memory. Ask Claude: "Which action items from this week's meetings are still waiting on me?" Granola remembers every meeting you've had, so your recaps quietly become your to-do list. Don't use Granola? Paste your own notes into the second prompt instead, the voice profile and the draft-in-Gmail steps work the same.
Level Up: Build the meeting-recap Skill If you do this after every call, encode it. Tell Claude: "Build me a skill called 'meeting-recap' that pulls my last Granola meeting and drafts the follow-up in my voice: decisions, action items with owners, one next step, kept short, saved as a Gmail draft to the attendees. Use my saved voice profile." Then "run meeting-recap" is the whole prompt. New to skills? See Claude Advanced, Section 1.

Method from Ruben Hassid's meeting-email framework (How to AI) • Granola is a third-party tool (granola.ai) • structured by Steve Gustafson

When Automations Go Wrong

Every workflow above teaches you how to build something. Almost nothing online teaches you what to do when one quietly breaks. Three failure modes cover most of it. Here is each, and the fix.

1. Dead input The source folder empties, a connector logs out, an export stops landing. The task keeps firing and produces nothing, or worse, a confident summary of nothing. Fix: a freshness check at the top of every task. Confirm the input exists and has new data. If not, report that and stop.
2. Stale repetition A task with no memory re-reports the same items every run. Day 20 looks like day 1. Fix: a state file. The task reads what it reported last time, does not repeat it, and logs what broke so the next run is smarter. This is self-healing, and it is the gap between a task that runs and a task that improves.
3. Unread output The task fires on schedule, the email lands, nobody opens it. Delivery is not the goal. Reading is. Fix: track whether you actually read the output. Anything unread two weeks running earns a demotion or a delete. An automation that produces unread work is not a win, it is noise.
The three-question test Before you trust any recurring task, ask three things. Does it check its own inputs? Does it remember its last run? Do you actually read what it sends? If any answer is no, that is where it will fail.

Inspired by Nick Saraev, Maker School

Pick One. Start Today.

You don't need all seven. Pick the workflow closest to your daily pain. Set it up. Run it for a week. Then come back for the next one.