learn-claude-with-phoebe / Session 2 of 6
Learn Claude with Phoebe · Session 2 of 6

Your AI chief of staff: daily automation

Morning briefings from your real calendar, critical email alerts that interrupt only when they should, one-sentence meeting booking in Slack, and a knowledge base that never forgets - all on our Claude Teams plan, and you build the core yourself today.

🟢 Easy Senior management C-level EAs & chiefs of staff 45 min live + self-study
0-3 · Welcome 3-18 · Concepts 18-40 · Build your own 40-45 · Q&A
Part 0

Why this session exists

Your mornings start with 40 unread emails, a packed calendar, and no time to figure out which three things actually matter. This session builds you an executive chief of staff on the tools we already pay for: Slack stays your front door, Claude Teams is the brain, and Google Workspace is where your data already lives. No third-party middleware, no new vendors.

Live - presented in session Self-study - read after class ★ Build-along demo 3 official courses covered
★ What you walk out with today Your own "Chief of Staff" Claude Project running a real morning briefing on YOUR calendar and inbox, a scheduled 8am auto-briefing, and a one-page plan (plus a ready-to-send request to IT) for the Slack and alert automations. Bring your laptop - the middle 22 minutes are you building.
Part 1 · the chief-of-staff pattern

One brain, two doors, six jobs 6 min live

Everything today follows one architecture: you talk to Claude through Slack or the Claude app, Claude reads your Workspace through governed connectors, and only summaries come back out. Six jobs, one pattern.

Your front doors 💬 Slack (@Claude / DM) 🖥️ Claude app + web 📱 Claude mobile Claude Teams Chief of Staff Project Your instructions + VIP list Scheduled 8am briefing Claude Tag in Slack Google Workspace 📧 Gmail · 📅 Calendar 📁 Drive · 🚪 Rooms via approved connectors Data flows right to left as summaries only: Claude reads what YOU can read, and Slack sees the digest, never raw mailboxes. Replaces the Gemini + Apps Script "Gem" pattern - same jobs, fewer moving parts, no API keys for the core.
🔍 Click to zoom - the chief-of-staff architecture on our stack
LiveThe six jobs your chief of staff does2 min
JobYou say / it happensBuilt in
1 · Morning briefing + to-do"Morning briefing" → today's meetings, conflicts, email triage, 3-5 item to-doDemo 1 + 2
2 · Critical email alertsEmails scored 1-5; only 4-5 (Chairman, budgets, deadlines) surfaceDemo 1 · Tier 3 for push
3 · Document search + knowledge base"Find the latest file on X" → doc found, 3-5 takeaways summarizedDemo 1
4 · Text-to-calendar"Book coffee with HR tomorrow 2pm" in Slack → event appears, bot confirmsTier 2
5 · Room & attendee coordination"Is Room A free tomorrow afternoon?" → 2-3 slots where rooms AND people are freeTier 2/3
6 · Auto briefing at 8amThe briefing is waiting before you open your laptopDemo 2
Real world

A managing director's Tuesday, after setup: 8:00 - the briefing is already in Slack: 6 meetings, one conflict flagged, two critical emails (a board member on budget, a contract deadline), a 4-item to-do. 8:05 - she DMs "@Claude move my 3pm to Thursday and book Room B" - done. 10:40 - an alert interrupts: the auditor needs a response by noon. Everything else waited for lunch. That is the chief-of-staff experience - roughly 30 minutes of triage handed back, every day.

LiveSecurity first - the slide for your own boss2 min

Lead with this before anyone asks. The whole system runs on tools we already license, inside our corporate perimeter:

  • No training on our data. Claude Teams and Enterprise data is never used to train models - contractual, not a setting.
  • Permission inheritance. Connectors act as YOU: Claude can only see the emails, events and files your own account can see. A briefing never leaks someone else's inbox.
  • Slack sees summaries only. The raw mailbox is read inside the Claude-to-Workspace connection; what lands in Slack is the processed digest.
  • No third-party middleware. No Zapier, no Make.com - direct, admin-governed integrations. Our sys admin can see, approve, and revoke every connector.
  • Human in the loop stays. The chief of staff drafts, flags, and proposes; sending, deciding, and signing stay with you (that is Diligence from Session 1).
Say it in one line "It runs on our existing Claude Teams and Google Workspace licenses, reads only what I can already read, never trains on our data, and Slack only ever sees the summary."
Self-studyWhat needs sys-admin approval - and the ready-to-send request2 min read

Three approvals unlock everything in this session. Copy the request below, adjust, send. (The full template also lives in the repo under materials/templates/.)

ApprovalUnlocksEffort for IT
Google Workspace connectors (Gmail, Calendar, Drive)Briefings, triage, doc search - the coreEnable in Claude admin console, one time
Claude app for Slack (Claude Tag)@Claude in channels, DMs, one-sentence bookingInstall from Slack Marketplace + workspace approval
Claude API key (optional, Tier 3 only)The push automations: 5-minute email watcher, room engineCreate a scoped key, store in Apps Script properties
★ Send to your sys adminSubject: Enable Claude connectors for the exec team - 3 items Hi [admin name], Following the Claude onboarding session, [N] of us on the senior team want to run the "chief of staff" workflow (morning briefing, email triage, document search) on our existing Claude Teams plan. Three requests: 1. Enable the Google Workspace connectors (Gmail, Calendar, Drive) for our Claude Teams workspace - they respect existing user permissions and are covered by the Teams no-training guarantee. 2. Install the Claude app for Slack from the Slack Marketplace and approve it for our workspace, so we can DM @Claude and use it in a private #cos-alerts channel. 3. (Optional, for the automation phase) Create one Claude API key for the automation service account, to be stored in Google Apps Script's Properties Service - used only for scheduled email scoring and room booking. Security posture: no third-party middleware, connectors inherit user permissions, Teams data is excluded from model training. Happy to walk through it - it took 45 minutes to learn, it'll take you less. Thanks, [name]
Part 2 · how it's built

Three tiers, from ten minutes to full autopilot 5 min live

You don't need everything on day one. Tier 1 you build yourself in the next 22 minutes. Tier 2 is one admin install away. Tier 3 is a one-time, ~90-minute job for one IT champion - and every executive benefits from the same engine.

Tier 1 · today, yourself Chief of Staff project + connectors: briefing, triage, doc search on demand Tier 2 · this week, admin Claude in Slack (@Claude, DMs), scheduled 8am auto-briefing, one-sentence meeting booking Tier 3 · IT champion, once Apps Script + Claude API engine: 5-min email watcher → #cos-alerts, room + attendee FreeBusy booking, push briefings to the channel, same code reused per executive No code · 10 minutes No code · 1 admin install ~200 lines · one time, everyone benefits
🔍 Click to zoom - the three build tiers
LiveTier 1 - the Chief of Staff project2 min · then Demo 1

A Claude Project (Session 1's "knowledge that compounds") with two things inside: connectors switched on (Gmail, Calendar, Drive) and standing instructions that turn generic Claude into YOUR chief of staff - who your VIPs are, what counts as critical, how you like your briefing. Once saved, every conversation in the project starts already knowing the rules. You build this yourself in Demo 1 with the exact instruction prompt below.

Real world

The instruction block is where the magic personalizes: a CFO lists the audit committee and "anything with the word covenant" as auto-critical; a CMO lists agency partners and campaign-budget threads. Same template, different VIP list - that one paragraph is the difference between a generic assistant and YOUR chief of staff.

LiveTier 2 - Claude in Slack + the 8am schedule3 min · then Demos 2-3

Slack stays your chat tool - Claude moves in. With the Claude app for Slack (Claude Tag, rolling out to Teams and Enterprise plans since June 2026) installed by your admin:

  • DM @Claude like a colleague: "Book a 30-min sync with the DS team tomorrow at 2pm", "Any urgent emails?", "Find the Q3 roadmap".
  • Tag @Claude in a channel - one shared Claude per channel that the whole team sees and can redirect mid-task. Perfect for a private #cos-alerts channel with you and your EA.
  • Admin-controlled: channel-level access controls; Claude in Slack respects the same workspace governance as everything else.

The 8am auto-briefing needs no code either: Claude's scheduled tasks run your briefing prompt on a timer - every weekday morning, the digest is waiting before you open your laptop. You configure yours in Demo 2.

End-to-end config (what your admin does once)

Slack Marketplace → search "Claude" → Add to Slack → approve for the workspace (workspace admin).

Connect the Claude workspace when prompted, so Slack conversations use our Teams plan (and its no-training guarantee).

Create the private channel #cos-alerts → invite yourself, your EA, and /invite @Claude.

Test: DM @Claude "hello" - then try "What's on my calendar tomorrow?" once connectors are approved.

Self-studyTier 3 - the push-automation engine (for your IT champion)3 min read

Two jobs genuinely need a scheduled engine rather than a chat: the 5-minute email watcher (score every new email 1-5 with Claude, push only 4-5 to #cos-alerts) and room + attendee FreeBusy booking across resource calendars. The original chief-of-staff pattern used Google Apps Script + the Gemini API - we keep the exact same architecture and swap in the Claude API. One file, ~200 lines, no servers.

The one function that changes// Apps Script: Claude replaces Gemini - everything else stays identical const ANTHROPIC_API_KEY = PropertiesService.getScriptProperties().getProperty('ANTHROPIC_API_KEY'); function askClaude(prompt) { const resp = UrlFetchApp.fetch('https://api.anthropic.com/v1/messages', { method: 'post', contentType: 'application/json', headers: { 'x-api-key': ANTHROPIC_API_KEY, 'anthropic-version': '2023-06-01' }, payload: JSON.stringify({ model: 'claude-sonnet-5', max_tokens: 1024, messages: [{ role: 'user', content: prompt }] }), muteHttpExceptions: true }); const json = JSON.parse(resp.getContentText()); return (json.content && json.content[0]) ? json.content[0].text : 'Could not process that request.'; }
  • Email watcher: Apps Script timer (every 5 min) reads unread Gmail → askClaude scores 1-5 with your VIP rubric → score ≥4 posts to the #cos-alerts webhook → thread labeled processed. Low-priority mail stays silent.
  • Room engine: parses "find a room for a 1-hour meeting with john@ Thursday morning" → FreeBusy query across room calendars AND attendees → returns 2-3 slots where everyone and a room are free.
  • Secrets hygiene: API key and Slack webhook live in Apps Script's Properties Service - never hard-coded, never in Slack messages.
  • The complete adapted script (config, Slack webhook helper, email watcher, room engine, daily briefing trigger, deploy steps) is in the repo: materials/templates/apps-script-claude-cos.md. Setup is ~90 minutes once; each additional executive is a config entry.
Build order that works Run Tiers 1-2 for two weeks first. Most executives find the scheduled briefing + on-demand triage covers 80% of the value - then Tier 3 is a deliberate upgrade, not day-one complexity.
Part 3 · the six jobs, properly

Prompts and patterns for each job 4 min live

Every job from the architecture, with the working prompt. In the session we do jobs 1-3 live inside your new project; the rest are configured in Tier 2/3.

LiveJob 1 · The morning briefing and to-do list1.5 min

The crown jewel. One word - "Morning briefing" - and your chief of staff scans today's calendar, triages 24 hours of email against your VIP rubric, and hands you a 3-5 item to-do list. The intelligence lives in the project instructions (Demo 1); the schedule makes it automatic (Demo 2).

Real world

What good looks like: "3 back-to-backs 9-11 with no prep gap - I moved your reading time. Conflict: the vendor call overlaps the leadership sync. Critical: board member Y asked for the revised budget by EOD (email 2h ago). To-do: 1) approve budget note, 2) 5-min prep for the 2pm negotiation - talking points attached, 3) sign the contract renewal before Thursday."

LiveJob 2 · Critical email alerts - the 1-5 triage funnel1.5 min
every incoming email (100%) scored 1-5 by Claude against YOUR rubric score 4-5 → alert (~5%) 5 = crisis (Chairman/Board, legal, financial) · 4 = high (MD/HoD, budgets, deadlines) 3 = normal, waits for your check · 2 = FYI · 1 = noise. Nothing below 4 interrupts you. You calibrate the rubric weekly.
🔍 Click to zoom - the triage funnel: only the top of the funnel interrupts
★ On-demand check (works after Demo 1)Check my priority emails from the last 12 hours. Score each 1-5 using my rubric (5=Crisis, 4=High, 3=Normal, 2=FYI, 1=Noise). Show ONLY score 4-5, each as: Sender - Subject - why it's critical - the action I must take. If nothing scores 4+, just say so - do not promote a 3.

The push version (alerts arriving in #cos-alerts within 5 minutes, unprompted) is the Tier 3 watcher - same rubric, on a timer.

LiveJob 3 · Document search and the team knowledge base1 min

Two complementary moves:

  • Search what exists: with the Drive connector, "Find the latest file on [project] and summarize the 3-5 core takeaways" replaces ten minutes of Drive archaeology - and it only finds files you already have access to.
  • Curate what matters: upload the 5-10 documents you're tired of re-explaining (strategy note, org chart, glossary, key policies) into the project itself. Now every briefing and draft is grounded in the same canon - that is your team knowledge base, and on Teams you can share the whole project.
★ Try in Demo 1Find the latest file on [a real project name] in my Drive. Give me: the file (linked), when it was last updated, and the 3-5 takeaways an executive needs. If several candidates exist, pick the most recent and list the others in one line.
Self-studyJobs 4-5 · Text-to-calendar and room coordination1.5 min read

Text-to-calendar: with the Calendar connector, booking is one sentence to @Claude in Slack or in your project: "Book a 30-min sync with [name] tomorrow at 2pm." Claude extracts title, time, attendees, confirms with you, creates the event. The chief-of-staff instructions make it always confirm before creating - Diligence by design.

Rooms and attendees: "Is Room A free tomorrow afternoon?" or "Find a slot Thursday morning where John, Sarah and a meeting room are all free." Simple checks work through the Calendar connector; the exhaustive cross-referencing version (every room × every attendee × 30-minute grid) is the Tier 3 FreeBusy engine, which returns "Option 1: 09:30-10:30 in Room A or Room C - all attendees free."

Real world

The showstopper moment from the original playbook stays the showstopper here: an MD types one sentence in Slack, the confirmation appears in the thread, and the event is on the calendar before the meeting about meetings would have even been scheduled.

Self-studyJob 6 + file management · Claude Cowork, your AI that works IN your files2.5 min read

Everything so far reads and summarizes. Claude Cowork (in the Claude desktop app) goes further: you point it at a folder and it works - renames and organizes files, drafts documents in place, processes a stack of PDFs, builds the deck skeleton. It runs a visible task loop:

1 Brief it 2 It plans 3 It works 4 You review "Organize this folder of 80 contracts by vendor and year" shows you the plan before touching a file renames, moves, drafts - visibly, step by step approve, correct, or roll back
🔍 Click to zoom - the Cowork task loop: it plans before it touches anything
  • Standing context: global instructions + projects work here too - Cowork picks up the same chief-of-staff rules.
  • Skills and plugins: teach Cowork your recurring file jobs ("month-end board pack assembly") once, reuse forever - that is Session 3's entire topic.
  • Claude in Chrome and Microsoft 365: the same assistant inside the browser and Office surfaces, so "where you work" stops mattering.
  • Safety practices: start with a copy of the folder, review the plan before approving, validate third-party skills before installing - the same Diligence habits, applied to files.
Demo 1 of 3

Build YOUR chief of staff ★ 12 min · everyone builds

By the end of this demo, a real morning briefing runs on your real calendar and inbox. Laptops open.

In Claude, create a new Project. Name it Chief of Staff.

Open Settings → Connectors → enable Gmail, Google Calendar, Google Drive and authorize with your corporate account. (If the toggles are missing, the admin approval hasn't landed - pair up with a neighbor who has it, and send the Part 1 request after the session.)

Paste the instruction prompt below into the project's instructions, then personalize the VIP list - replace the role placeholders with the real names that matter to you.

Type: "Morning briefing". Watch it scan YOUR day. Then run: "Check priority emails from the last 12 hours" and "Find the latest file on [a real project]".

Discernment lap (Session 1 habit): tell it one thing to fix - "too long, cap the briefing at 150 words" or "add my 1:1s' prep notes". The instruction learns when you re-save it.

★ The chief-of-staff instructions - paste into your projectYou are my executive chief of staff. You have access to my Gmail, Google Calendar, and Google Drive through connectors - use them. 1. MORNING BRIEFING - when I say "Morning briefing": - Scan today's calendar: bulleted meeting list (times, attendees), flag conflicts and missing prep gaps. - Scan the last 24h of email. Flag CRITICAL ALERTS using my rubric: * SENDER: Chairman, board members, managing directors, heads of department, my assistant. [REPLACE WITH REAL NAMES] * TOPIC: budget, finance, invoices, contracts, audit, compliance, urgent HR, risk escalations. [ADJUST TO YOUR WORLD] * TIMING: anything needing a response today or within hours. - End with a crisp 3-5 item to-do list, most critical first. 2. PRIORITY EMAIL CHECK - when I say "Check priority emails": - Score each email 1-5: 5=Crisis, 4=High, 3=Normal, 2=FYI, 1=Noise. - Show ONLY 4-5: Sender - Subject - why critical - action required. Never promote a 3. 3. DOCUMENT SEARCH - when I say "Find a file on [topic]": - Search my Drive, return the best match linked, last-updated date, and 3-5 executive takeaways. 4. MEETINGS - when I ask to schedule or check availability: - Check attendees' and rooms' free/busy where you can, propose 2-3 slots, and ALWAYS confirm with me before creating any event. Tone: professional, direct, executive-ready. Lead with critical items. Bullets, no fluff. If data is unavailable (connector off, no results), say so plainly - never invent.
Data tip Connectors read your real inbox - so today's practice IS production. Nothing needs pasting anywhere, which is exactly the point: the data never leaves the governed connection.
Demo 2 of 3

Schedule the 8am auto-briefing ★ 5 min · everyone configures

The difference between a tool and a chief of staff: the chief of staff shows up unasked. Two minutes of configuration, then it's every weekday, forever.

In Claude, ask: "Every weekday at 8:00am, run my morning briefing from the Chief of Staff project and send it to me." Claude sets up the scheduled task; confirm the schedule it proposes.

Check the schedule: ask "list my scheduled tasks" - you should see the weekday 8am briefing. Adjust the time to beat your commute.

Tomorrow morning: read the briefing that arrived before you did. Note ONE thing it missed or over-flagged - that's your calibration input for the week.

Real world

Week one, the briefing flags too much - everyone's rubric starts too broad. The calibration habit from the original playbook applies unchanged: each day ask "did the briefing miss anything? did it cry wolf?" and adjust one line of the instructions. By week two, the signal is clean, and that's when executives stop checking their inbox before breakfast.

Demo 3 of 3

Slack: one sentence, meeting booked ★ 5 min · follow along

If the Claude app for Slack is already installed, everyone does this live; if not, watch the presenter and this becomes item 2 on your admin request.

In Slack, DM @Claude: "hello" - meet your chief of staff's front door.

The showstopper: "Book a 30-min sync with [a real colleague] tomorrow at 2pm." Confirm when it asks. Open Google Calendar - the event is there, booked from one sentence.

Try: "Is Room A free tomorrow afternoon?" and "Any urgent emails in the last 4 hours?"

In your team channel, tag @Claude on a real thread: "@Claude summarize this thread and list the action items with owners." One shared Claude, visible to everyone - notice a colleague can redirect it mid-task.

Why Slack and not just the Claude app? Because it's where you already are at 8:03am. The #cos-alerts channel pattern keeps everything auditable in one place: scheduled briefings, critical alerts, and your one-off requests - and your EA sees the same picture you do.
After the session

This week, and how to ask follow-up questions ◐ 30 min total

Source material

Official courses and playbooks covered

This session teaches the working content of three official courses from claude.com/resources/courses, plus our internal Chief-of-Staff setup playbook, adapted from its original Gemini + Apps Script design to Claude Teams. The 4D-framework halves of the two AI Fluency courses were taught in Session 1 - here you applied them.

Introduction to Claude Coworktask loop, standing context, skills & plugins, Chrome/M365, safety - Part 3 + Session 3 continues skills
AI fluency for small businesses0.9 hr · refine loop, transparent AI use, human in the loop - applied throughout; 4D core in Session 1
AI fluency for nonprofits54 min · researching/writing with AI, privacy & data, workflow augmentation - applied throughout; 4D core in Session 1
Chief of Staff AI setup guide (internal)all 7 phases: gem→project, Slack app→Claude Tag, bot brain→connectors, calendar, alerts, rooms, auto-briefing
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Why can your chief of staff only see YOUR inbox, never a colleague's?

Permission inheritance is the security model - governance by construction, not by promise. That's the line for your own boss.

2 · Which emails are allowed to interrupt you?

The triage funnel: ~5% interrupts, everything else waits for your check. Calibrate the rubric weekly until the signal is clean.

3 · What does Tier 1 (build it today, yourself) require?

The core (briefing, triage, doc search) is a 10-minute no-code build. Code only enters at Tier 3, once, for the push automations.

Session 2 cheat sheet · pin this

The architectureSlack + Claude app in front · Claude Teams project as the brain · Workspace connectors behind. Summaries out, never raw data.
Morning briefing"Morning briefing" → schedule it for 8am weekdays. Calibrate one line daily for two weeks.
Email triageScore 1-5 against YOUR VIP rubric. Only 4-5 interrupt. Never let it promote a 3.
Booking line"Book a 30-min sync with [name] tomorrow 2pm" - to @Claude in Slack. Always confirm before create.
Knowledge baseUpload your 5 most re-explained docs to the project. Share the project with the team.
Security one-liner"Runs on our licenses, reads only what I can read, never trains on our data, Slack sees summaries."
Three tiers1: project (you, today) · 2: Slack + schedule (admin, this week) · 3: push engine (IT, ~90 min once).
Follow-up pathAsk Claude itself → #learn-claude channel (@Claude answers first) → next session's Q&A.