PRESENTS · MAVEN COHORT 2 · SESSION 2 OF 5 · WEEK 3 SAT MAY 30, 2026 · 12:00 PM EDT How to Build AI. Three engine types · three tool worlds · one Chief of Staff you'll ship today. Module 2 · Week 3 · Live on Maven CO-INSTRUCTORS S Sandhiya Thiruvengadam Leader, HRBP & AI · Apexon Q Q Hamirani Chief People Officer, Highlevel · Adjunct Faculty, LBS Session 2 · Slide 1/17
FROM YOUR WEEK 1 FEEDBACK We heard you. Today we go slower. Fewer concepts per beat. Two hours together · so we have room to breathe. We don't have to cover everything. DM us anytime. Book a 1:1. We're with you on this journey. ↩ HOME Session 2 · Slide 2/17
Last week · what you created Two things landed across the room. ★ ONE You started a Skill. A reusable AI pattern of your own. Something Claude can run on demand in the shape that matches how you work. EXAMPLES FROM THE ROOM "JD writer · Senior PM, 5 IC tiers." "Compensation outlier flagger." "Tone re-writer for tough conversations." "PIP drafter that doesn't sound mean." ★ TWO You mapped your quick wins. Your repetitive tasks plotted on a 2x2 grid. Impact on one axis · complexity on the other. Each of you now has a starting list. WHAT THE DATA + GUT SAID JD writing & intake notes dominated Quick Wins. BGV + onboarding sat in Strategic. Comp + perf review questions surfaced too. Pattern: most of you can ship something this week. Today we turn those two things into one working agent. ↩ HOME Session 2 · Slide 3/17
How to use Claude · the restaurant Same map from Session 1. Today we sit at the table. Your homework brain becomes your regular. FRONT OF HOUSE · what the guest sees ✓ SESSION 1 ① WALK IN Chat claude.ai · type back-and-forth Covered last week. You should already have 1 Skill installed. YOU ARE HERE ② SIT AT A TABLE Claude Project a contained workspace with memory One Project per recurring task. Where your CoS lives. YOU ARE HERE ③ THE FILE ON YOU CLAUDE.md the profile that personalizes Your role, voice, rules · loaded every session · seeds your CoS. ✓ SESSION 1 ④ THE MENU Skills markdown packs Claude can call Anatomy from Week 1 Slide 12. We add to your CoS today. ⑤ STANDING ORDERS Routines scheduled actions Claude runs for you "Every weekday 7AM: brief me." Built in Session 3. ⑥ THE RUNNER Dispatch parallel sub-agents on one job "Research 5 candidates at once." Met in Session 4. ⑦ KITCHEN LINE Computer Use Claude drives a browser for you "Pull these 20 PDFs and tag them." Met in Session 4. ⑧ THE RECEIPT Hooks what fires after every meal "After every reply, log to Notion." Mentioned in Session 4. ⑨ BACK OF HOUSE · the supply chain (drawn large because it carries everything) APIs · Connectors · MCP Today (Slide 9) we meet Connectors. The full Trio is unpacked in Session 4. Granola, Slack, Drive · each gives Claude one new sense. API Connector MCP Today we touch Connectors. MCP comes Session 4. DOORS ↔ ZONES Door 1 ✓ (Chat covered last week). Door 2 = TODAY (Table + Regular + Menu). Door 3 = sessions ahead. Your homework (Project 1.2) is the brain you'll wire into Project + CoS today. ↩ HOME Session 2 · Slide 4/17
The three categories of agents A ladder · each step gives the agent more freedom to decide things on its own. Category 1 Deterministic automation ★ TODAY'S FOCUS Category 2 Reasoning and acting agents Category 3 Multi-agent networks REFERENCE · LENNY'S NEWSLETTER · AGENT TAXONOMY ↩ HOME Session 2 · Slide 5/17
CATEGORY 1 Deterministic automation Fixed input · fixed trigger · fixed output. The LLM, if it shows up at all, lives in a small box. 1 Email rule PURE DETERMINISTIC · NO LLM Input: An email arrives. Trigger: Body contains "[INVOICE]". Output: Move to /Finance folder. Rule-based. Predictable. No AI in the loop. 2 Email + LLM step LLM IN A BOX · ONE JOB Input: Email from your CFO. Trigger: Sender match. LLM step: Summarize · classify intent · draft 2-line reply. Flow is fixed. LLM is in a box doing one small job. 3 BGV trigger REAL HR WORKFLOW Trigger: Offer status changes to "accepted" in Excel · BGV link pasted. 10–15 min: Auto-email · "Congrats · here's your BGV link · please complete." 24 hr later: If incomplete · reminder email. Light LLM drafts wording. Path stays locked. ★ TAKEAWAY Most HR automations live here. Start here · ship something small · then graduate. ↩ HOME Session 2 · Slide 6/17
CATEGORY 2 ★ TODAY'S FOCUS Reasoning + acting agents The dynamic path. You provide the goal and the tools · the agent figures out the how. ★ THE LOOP Instead of scripting step 1, step 2, step 3 ... you state your goal. The agent reasons about the situation, picks a tool from the ones you gave it, acts, assesses what happened · and loops. THE LLM IS AT EVERY STEP not just in the middle. GOAL what you want done Reason Act Assess tool tool tool tool tool tool ★ TODAY'S FOCUS The Chief of Staff Agent Sandhiya runs hers in MS Copilot. ★ THE TOOLS email read · email send calendar peek · tag Excel lookup · update Slack post · Slack search you give them · agent picks. Today = Category 2. One agent. Reason · Act · Assess. Loop. ↩ HOME Session 2 · Slide 7/17
CATEGORY 3 Multi-agent networks Several agents working together · one hands off to another · a team of specialists. ★ WORKED EXAMPLE · THREE AGENTS, WORKING TOGETHER ORCHESTRATOR Chief of Staff agent your daily brief SPECIALIST What People Owe Me open asks · follow-ups SPECIALIST Transcript to MOM meeting minutes Owe-Me reads from MOM. CoS asks Owe-Me. Your morning brief surfaces blocks across all three. We're naming Category 3 · not entering it today. Walk before we run. ↩ HOME Session 2 · Slide 8/17
CALLBACK · WEEK 1 · THREE DOORS Today's setup · three doors, three worlds. Same Three Doors language as Week 1 · different angle today. Week 1 was about levels of AI usage. Today is about tool ecosystems where agents live. DOOR 1 Microsoft Copilot IF YOU LIVE IN Outlook · Teams · Excel · SharePoint ★ CATEGORY 1–2 · INTEGRATED DOOR 2 Google Automation IF YOU LIVE IN Gmail · Sheets · Drive · Apps Script ★ CATEGORY 1–2 · WORKFLOWS DOOR 3 Claude Routines + CoS IF YOU LIVE IN Claude Project · Routines · MCP ★ CATEGORY 2 · TODAY'S FOCUS Watch one example agent in each. Then pick yours. ↩ HOME Session 2 · Slide 9/17
DOOR 1 LEVEL 1 · PERSISTENT HELPER Microsoft Copilot Agent Persistent helper · build it once, your team uses it. ★ WHAT YOU'RE LOOKING AT Copilot Agent Studio. Instructions · what the agent should do. Knowledge · SharePoint, Teams, Outlook plug in here. Capabilities · what it can do (write, search, act). TAKEAWAY If your company runs on Microsoft 365 · this is where your CoS lives. ↩ HOME Session 2 · Slide 10/17
DOOR 2 LEVEL 1 · PERSISTENT HELPER Google Gemini Gem Persistent helper · paste prompt, drop files, run. ★ WHAT YOU'RE LOOKING AT The simplest build of the three. Name + Description. Instructions · the system prompt that shapes voice. Knowledge · drop in files the Gem can reference. Run · talk to it in Gemini. TAKEAWAY If your company runs on Google Workspace · this is your stack. ↩ HOME Session 2 · Slide 11/17
DOOR 3 TWO RUNGS · SAME PLATFORM Claude · two levels of autonomy Start at Level 1 · graduate to Level 2 when you're ready to let it run without you. LEVEL 1 Claude Project Persistent helper · you drive each turn. Instructions + Files. You ask · it answers · context sticks. LEVEL 2 Cowork Scheduled Task Autonomous run · fires without you · picks tools · decides. Same prompt · runs daily at 9 AM · posts to your Slack. Same platform · two rungs of the ladder. ↩ HOME Session 2 · Slide 12/17
SCALE · the 5 pillars How to evaluate any AI tool you build (or buy) for HR. Same framework, vendor or DIY. S Security Where does data live? SOC 2? GDPR? Audit trail? C Cost Per-seat? Per-token? ROI vs. status quo? Hidden infra cost? A Accuracy Hallucination rate? Bias on HR data? Quality at scale? L Learning Improves with feedback? Or stays static? Custom training? E Ease Setup < 1 hour? Onboarding curve? Integration friction? ★ DEEP-DIVE RESOURCE · IN YOUR PACK THIS WEEK (SLIDE 11) HR Tech Vendor Eval · 60+ questions sorted by SCALE pillar. Lift from our Cohort 1 library. Paste vendor responses · score 1–5 · use it in your next AI tool pitch. file: HR_Tech_Vendor_Eval_SCALE.csv · download from Maven student home Run your own CoS through SCALE too. If you'd reject a vendor for it, fix it in your build. ↩ HOME Session 2 · Slide 13/17
Connectors · plug your CoS into the world A connector is how your assistant stops asking and starts knowing. YOUR CoS · MCP HUB YOUR CoS # Slack posts, threads, DMs M Gmail inbox + drafts 31 Calendar events, free/busy G Granola meeting notes N Notion docs, wikis ALL CONNECTED VIA MCP · OPEN PROTOCOL REAL-WORLD ROUTINE · MY DAILY LEARNING Granola → daily summary → Slack 1 CAPTURE Granola transcribes every meeting 2 SYNTHESIZE CoS pulls one learning per day 3 DELIVER Slack DM lands on my phone at 7 AM CoS My CoS 7:02 AM ★ One thing from yesterday: When the CHRO said "we'll circle back" in Tuesday's meeting, she meant "I don't want to decide yet." Slow down the comp ask. from 3 meetings · total reading time 24 seconds From a routine I posted on LinkedIn · Jan 2026 · #genaiinmyplaybook Without connectors, you have a chatbot. With them, you have an assistant. ↩ HOME Session 2 · Slide 14/17
Anatomy of a Chief of Staff prompt Sandhiya's actual MS Copilot template. 6 sections. Same skeleton in Claude. This is Project 2.1. 07_Shareable_Prompt_Template.md · Sandhiya's MS Copilot CoS 6 THINGS THIS PROMPT DOES 1 # ─── ROLE ───────────────────────────── 2 You are [YOUR NAME]'s Chief of Staff, 3 a sharp, honest, proactive AI. 5 # ─── PERSONALITY ────────────────────── 6 Direct. Doesn't sugarcoat. Warm but 7 efficient. Speaks like a trusted advisor. 9 # ─── TOOLS (connectors) ─────────────── 10 Get emails · Get calendar events 11 Get Teams messages · Find meeting times 12 Get tasks 14 # ─── MORNING BRIEFING (8 sections) ──── 15 1. Today's calendar 16 2. Email highlights (needs reply / FYI) 17 3. Teams highlights · 4. Action items 18 5. Pending with you · 6. Sent & waiting 19 7. Focus areas · 8. Suggestions 21 # ─── FORMATTING ─────────────────────── 22 Tables. Bold names. Emojis sparingly. 23 Scannable in under 3 minutes. 25 # ─── PRIORITIZATION ─────────────────── 26 1. Items from [TOP STAKEHOLDER] 27 2. Deadlines today or overdue 28 3. Items blocking others 29 4. Everything else by recency ⚙ ROLE · who Claude is being ⚙ PERSONALITY · how Claude speaks ⚙ TOOLS · what Claude can read/do ⚙ BRIEFING · 8 sections that fire daily ⚙ FORMATTING · how output looks ⚙ PRIORITIZATION · whose stuff first PORTS ACROSS PLATFORMS ↑ Sandhiya's CoS lives in MS Copilot Studio. Same 6-section recipe ports to Claude Projects, ChatGPT Custom GPT, anywhere. Project 2.1 = your port. ★ HOW THIS TRANSLATES TO YOUR STACK same idea, different brand · pick the door your company already pays for ★ CLAUDE · RECOMMENDED Claude Project · CoS system prompt MICROSOFT COPILOT Copilot Studio agent (Sandhiya's lives here) GOOGLE GEMINI Gem with custom persona OPENAI CHATGPT Custom GPT with instructions ↩ HOME Session 2 · Slide 15/17
🤖 Use AI Now · Build your CoS v1 7 minutes. Open Claude Projects. Follow these 5 steps. Paste the block on the right. 1 claude.ai → New Project Name it "My Chief of Staff." 2 Upload working-style doc Drop Project 1.2 into knowledge. 3 Paste the block → Project instructions panel. 4 Run 3 test queries Inbox triage · 1:1 prep · update. 5 3-Pass weakest output "Suggest 3 → incorporate." PASTE THIS · Project instructions panel 4 BLOCKS THIS DOES 1 You are my Chief of Staff. 2 3 # Read this first 4 Before any reply, read my 5 working-style doc in Project 6 knowledge. Match my tone. 7 8 # How to think 9 Think step-by-step. First 10 classify what I'm asking, 11 then reason, then answer. 12 13 # Default format 14 Bullets. Exec-summary up top. 15 Markdown tables when useful. 16 17 # Honest-day reader 18 If I'm replying late, working 19 weekends, or saying yes to 20 things I should say no to: 21 flag it. Don't sycophant. ⚙ READ FIRST · ground in your style ⚙ HOW TO THINK · Chain-of-Thought ⚙ FORMAT · how output looks ⚙ HONEST-DAY READER · flag overdrive ★ AFTER YOU RUN IT Drop a screenshot in cohort chat with #cos-v1. NEW WORDS ↑ CHAIN-OF-THOUGHT · HONEST-DAY READER Chain-of-Thought = "think step-by-step before answering." Honest-day reader = "flag when I'm overdoing it." Both make the output more honest. ↩ HOME Session 2 · Slide 16/17
Take these with you Resources on the left · actions in the middle · pick your rung at the bottom. ★ RESOURCE PACK · THIS WEEK You said "too many resources." This is the reply. ★ NEW THIS WEEK The AI Agency Ladder A deeper read of Lenny's 3-categories framework. Pair with today's session. Restaurant Map Week 1 callback · the whole Claude ecosystem. Skills Build Kit Week 1 callback · 20 HR skills you can build. CoS Template Today's anchor · fork it · ship your v1. ★ BUILDER PATH · DO THIS WEEK Three actions · not reading. ACTION 1 · 5 MIN Create a GitHub account Free. github.com/join. Sandhiya will walk through this live. ACTION 2 · 10 MIN Install Claude Code On your machine. claude.ai/code. We deepen in Session 3. ACTION 3 · BOOK IT Book a 1:1 with us Any slot before Session 3 · Calendly link in Slack. ★ BEFORE YOU GO · PICK YOUR RUNG · WHICH ONE WILL YOU SHIP BY WORKSHOP 3? CATEGORY 1 Deterministic automation Power Automate · Zapier flow · light LLM step. CATEGORY 2 Reasoning + acting agent Claude Project · Custom GPT · Gem · Copilot Agent. CATEGORY 3 Multi-agent network Walk before you run · revisit Session 4. Tell us in Slack which rung you're shipping · we'll pair you up. ↩ HOME Session 2 · Slide 17/17
End of Session 2. You shipped. Now keep the muscle warm. ★ NEXT SESSION Session 3 · Make it safe to ship Saturday June 6, 2026 · 12:00 PM EDT · live on Maven OPEN DECK → Between now and then · 10 minutes a day on Atlas 30. Ship one thing. ↩ HOME End of Session 2
SESSION 2 DECK · COHORT 2

Cohort 2 is in the room.Session 2 — Build the AI that works like you. These slides walk through it live with Q & Sandhiya. Cohort members sign in to open the deck; everyone else: catch Cohort 3.

ALREADY IN COHORT 2?

Welcome back. The deck loads behind the sign-in.

WATCHING FROM OUTSIDE?

Cohort 3 opens later this summer. We'll save you a seat.

CATCH COHORT 3 →
🏠 1 / 13