Mentor Playbook · Cohort 1

The session-by-session
runbook for Harvard mentors.

This is the same brief every mentor receives before Cohort 1. Minute-by-minute on Zoom — the cold open, the exact prompts, the breakout structures, the failure modes, the line you say when a fellow is stuck. Written by the Head of Pedagogy. Read it like a flight plan.

APPLY TO MENTOR See curriculum →
1:20 mentor ratio 120 min live, every Saturday 48 hr written-feedback rule 1 live capstone panel
Your job, in five sentences

You are not the lecturer.
You are the editor.

If you internalise these five sentences, you can run any session in this playbook. If you don't, no script will save you.

i.

You are not the lecturer — the lead analyst is. You are the editor. Your job is to make the fellow's artefact 30% sharper before midnight.

ii.

Your other job is to ask the question the fellow doesn't want. "What page?" "What kills this?" "Where did AI lie?" Never let it slide.

iii.

You hold 1:20 attention. You know each fellow's name, their capstone direction, and the last thing they got wrong. By Week 3 you cold-call by first name.

iv.

You grade the AI Workbook, not just the artefact. If a fellow can't show what AI gave them and what they overrode, it doesn't count.

v.

You write one feedback line per fellow within 48 hours. Format: kept · cut · push. No exceptions. This is the program.

The 120-minute clock · mentor cues

Where you talk.
Where you shut up.

The lead analyst owns the first 60. You own the second 60. But you're never passive — here's where you intervene.

00:00 → 00:15Hook
00:15 → 01:00Concept
01:00 → 01:45AI Lab
01:45 → 02:00Review
Hook · 15m

Camera on, mic off. Drop the cold-call name in chat at 00:03. Drop the poll at 00:08. At 00:12 unmute and ask the reveal question. Never read the hook — the lead does. You watch faces.

Concept · 45m

You are a metronome. Every ~12 min you drop one prompt in chat: a vocabulary check, a number to verify, a "what would break this?" At 00:45 you DM the 2 fellows who haven't spoken yet and ask them to take the next question.

AI Lab · 45m

Breakouts open at 01:02. You rotate through every room with a 4-min stopwatch. One question per room: "What did AI get wrong?" Pin the prompt starter and the 3 red-pen questions at 01:00. Pull stuck fellows into a side room.

Mentor review · 15m

Spotlight 2 fellows. 1-line praise, 1-line push, 1 standard question. End at 01:58 with the "kept · cut · push" preview. At 01:59 confirm the artefact deadline (Sunday 23:59 ET) in chat.

Zoom mechanics

The tools you use.
The tools you don't.

Default to the simplest mechanic that moves the artefact forward. If it doesn't make the artefact sharper, don't use it.

Breakouts

Pairs for IPS critique (W3) and 10-K verify (W4). Rooms of 4 for venture pitch (W5–W6). Never 6+. You rotate on a 4-min stopwatch.

Polls

One per session, max. W3 bias inventory poll runs anonymously. W7 scam-archetype poll runs publicly. Polls are diagnostic, not entertainment.

Whiteboard

W2 portfolio sketch only. Lead draws the 3-asset weights; fellows stress-test live. Disable for every other week — it eats time.

Spotlight

W8 capstone defence. Spotlight one fellow at a time, panel cameras visible bottom strip. Audience cameras off, mics off, chat moderated.

Chat rule

No AI-generated chat answers. If a fellow drops a paragraph that's clearly LLM output, DM them: "Paste your prompt + your edit, or it doesn't count."

Shared doc

The AI Workbook (a pinned Google Doc per fellow) is always open on a second screen. You sign each entry weekly with date + initials.

Session-by-session drill-down

Eight Saturdays.
Eight runbooks.

Each card is the full mentor brief for that session. Read it Friday night. Re-read the Hook block Saturday morning before login.

W1
Sat, Aug 29 · 9 AM ET
Research

Investing fundamentals

Learning outcome: Fellow can name the four asset classes, the three returns each produces, and one thing that kills each — without notes.

Pre-session prep · Friday

(1) Pull current 1-yr returns for VTI / BND / VNQ / GLD; screenshot ready in slide deck. (2) Re-read the SEC prospectus first page of VTI — you'll need a fact-check answer. (3) Open each fellow's intake form; note 3 fellows to cold-call by name. (4) DM the co-mentor with the breakout pairing list (alphabetical pairs, mixed grade level).

00:00 → 00:15
Hook · Cold open
  • 00:00 — Lead drops 4 tickers on screen, no labels. "Pick the winner 1995–2025."
  • 00:03 — You drop a poll: A / B / C / D. Wait 90 sec.
  • 00:06 — Reveal: no single winner. S&P wins 1990s, gold wins 2000s, real estate wins early 2010s.
  • 00:10 — You unmute. "If no one wins always, what does that tell us about owning just one?" Cold-call the 2 fellows who voted same as their neighbour.
  • 00:14 — Lead transitions: "Today we name what those four things actually are."
00:15 → 01:00
Concept · 3 teaching moves
  • Analogy: "An asset is a claim on future cash flows" → a vending machine you own a share of.
  • Board sketch: 2×2 grid — cash flow now vs. cash flow later, low vs. high risk. Drop each asset class in a quadrant.
  • Vocabulary anchor: ETF / mutual fund / index / expense ratio / yield / duration. Spell them in chat once each.
  • Cold-call beat: 00:38 — "Name one thing that kills a bond." Acceptable: rates rise, default, inflation.
01:00 → 01:45
AI Lab · Asset Profile Card
  • Solo work, no breakouts this week. Each fellow picks one ticker (VTI, BND, VNQ, GLD, or one stock).
  • Pinned prompt starter: "Explain [ticker] in one paragraph for a 10th grader: what it is, who it's for, and what kills it. Cite the prospectus or 10-K page where you got each claim."
  • Your 3 red-pen questions (DM to every fellow at 01:12): (a) Did AI cite a real page or guess? (b) What "kills it" did AI omit? (c) Rewrite the paragraph in your own voice in 4 sentences.
  • You rotate every 5 min into the highest-need DM. Pull stuck fellows into a 1:1 breakout for 90 sec.
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 fellow who picked a stock, 1 who picked an ETF.
  • Standard question: "What did AI tell you that the prospectus didn't say?"
  • 1-line praise + 1-line push: "Kept: your 'duration risk' line, that's correct. Push: cite the page next time — no page, no claim."
  • Failure mode to call out: the fellow who copy-pasted AI verbatim. Name it kindly, then re-prompt them in front of the room.
Artefact + gradingAsset Profile Card. Heaviest on Rigour (page cited) and Judgment (AI hallucination annotated). 7/10 = correct but uncited. 9/10 = correct, cited, with one annotated AI error in the fellow's own voice.
Common pitfalls(1) Fellow picks a meme stock and can't find a 10-K — redirect to ETF. (2) Fellow accepts AI's "low risk" claim on a junk bond ETF — ask them what the credit rating is.
Mentor 1-liner"What kills it? If you can't name what kills it, you don't own it yet."
W2
Sat, Sep 5 · 9 AM ET
Model

Portfolio thinking

Learning outcome: Fellow can build a 3-asset portfolio in a spreadsheet, run it through three regimes, and defend the weights in 60 seconds.

Pre-session prep · Friday

(1) Pre-load the 3-asset spreadsheet template in Google Sheets, share with view-only link. (2) Memorise the 2008 / 2020 / 2022 drawdown numbers cold. (3) Re-read each fellow's W1 Asset Profile Card — you'll quote one back to them. (4) Test the whiteboard tool; this is the only week you use it.

00:00 → 00:15
Hook · 100% vs 60/30/10
  • 00:00 — Lead shows two equity curves: 100% S&P vs. 60/30/10, 2000–2024.
  • 00:05 — Reveal average annual return is nearly identical. Reveal max drawdown is half.
  • 00:08 — You unmute: "Which one would let your mom sleep?"
  • 00:12 — Cold-call last week's quietest fellow by name. The point is one word: volatility.
00:15 → 01:00
Concept · 3 teaching moves
  • Analogy: Diversification is a band, not a solo. Same notes played by 3 instruments — one drops out, song continues.
  • Whiteboard sketch: draw correlation arrows. When does gold zig while stocks zag? Live audience input.
  • One equation on screen: real return = (Σ weight × return) − fees − inflation. Walk through 60/30/10 example slowly.
  • Cold-call beat: 00:50 — "Name one regime that breaks 100% bonds." Acceptable: 2022, 1970s stagflation.
01:00 → 01:45
AI Lab · 3-asset thesis
  • Solo build first (15 min). Pinned prompt starter: "Give me 10-yr historical CAGR and worst calendar year for VTI, BND, GLD. Cite source. If you don't know, say so."
  • Pairs breakout at 01:18 (8 rooms of 2). Each pair stress-tests the other's portfolio against 2008, 2022, and synthetic stagflation.
  • Your 3 red-pen questions: (a) Did your AI return numbers match the SEC filing? (b) Did you fee-adjust? (c) What's your worst calendar year — in dollars, not percent?
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 conservative portfolio (40/50/10) and 1 aggressive (90/5/5).
  • Standard question: "Defend your weights. Why not 80/20?"
  • Failure mode: fellow who can't name their worst calendar year — they didn't actually run the model.
Artefact + grading3-Asset Portfolio Thesis + stress table. Heaviest on Rigour. 9/10 = weights defended in writing, 3 regimes modelled, "what I'd change at 65" included.
Common pitfalls(1) Fellow builds 100% equities and calls it "aggressive" — ask what their drawdown tolerance is, in dollars. (2) AI hallucinates 10-yr CAGR — make them cross-check on the SEC site live.
Mentor 1-liner"In dollars, not percent. How much do you lose in your worst year?"
W3
Sat, Sep 12 · 9 AM ET
Verify

Risk, return & behaviour

Learning outcome: Fellow can name 6 biases, the blow-up that maps to each, and write a personal IPS that pre-commits against at least 3.

Pre-session prep · Friday

(1) Re-read the LTCM 8-min story — you'll narrate the second half live. (2) Build the anonymous bias-inventory poll (6 questions, multiple choice). (3) Pull each fellow's W2 stress table — you'll ask: "Given your worst year, what's your sell rule?" (4) Print the IPS template; one column for allocation, one for "I will sell if…".

00:00 → 00:15
Hook · LTCM
  • 00:00 — Lead opens with "Two Nobel laureates. $4.6 billion. Four months."
  • 00:06 — You take over the second half. Narrate the August 1998 Russia default beat-by-beat.
  • 00:11 — Drop the bias poll. Anonymous. Reveal the cohort's most common bias at 00:14.
  • 00:14 — "It wasn't bad math. It was bad behaviour. That's what we fix today."
00:15 → 01:00
Concept · 6 biases, 6 blow-ups
  • Anchoring → dot-com cost basis. Recency → GameStop. Confirmation → FTX retail. Loss aversion → 2008 mortgages. Overconfidence → Archegos. Herding → meme cycle.
  • Frame: Risk is not volatility. Risk = permanent loss of capital.
  • Cold-call beat: 00:42 — "Name the pre-commitment device that beats anchoring." Acceptable: automatic rebalancing, IPS with sell rule.
01:00 → 01:45
AI Lab · IPS v1 → v2
  • Solo draft 15 min: 1-page IPS — allocation, rebalancing rule, "I will sell if…" rule.
  • Pinned prompt starter: "You are a behavioural economist. Read this IPS and find the three biases hiding in it. Be specific — quote the line."
  • Pairs breakout at 01:25: swap IPS docs, role-play 2008. "Your portfolio is down 40%. Your friend just sold. What does your IPS make you do?"
  • Your 3 red-pen questions: (a) Is the sell rule a number or a feeling? (b) What's the rebalance cadence? (c) Did AI find a real bias or invent one?
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 strong IPS (numbered rules), 1 weak (vibes).
  • Standard question: "Read me the sentence in your IPS that would have stopped you from panic-selling in March 2020."
Artefact + gradingPersonal IPS v2. Heaviest on Judgment + Honesty. 9/10 = sell rule is numeric, at least 3 biases pre-committed against, AI critique annotated.
Common pitfalls(1) "Sell rule" is "if I feel scared" — reject, ask for a number. (2) Fellow lets AI write the whole IPS — the AI Workbook will catch this; flag in feedback.
Mentor 1-liner"Is that a number or a feeling? Numbers go in the IPS."
W4
Sat, Sep 19 · 9 AM ET
Research

Reading a business

Learning outcome: Fellow can pull 5 numbers from a real 10-K, cite the page, and write a 1-paragraph health verdict that survives "where on the page?"

Pre-session prep · Friday

(1) Download two anonymised 10-K PDFs — Costco and a near-bankrupt retailer. Redact names & logos. (2) Bookmark the cash-flow page on each. (3) Each fellow has pre-picked a public company in their intake; load the latest 10-K PDF for each in tabs. (4) Have the SEC EDGAR URL ready to paste.

00:00 → 00:15
Hook · 2 anonymous 10-Ks
  • 00:00 — Drop both PDFs in chat. "You're a bank. Lend $10M to one. 90 seconds."
  • 00:03 — Poll: A or B. Don't reveal yet.
  • 00:08 — Cold-call the fellow who picked the bankrupt one. "What number did you trust?" Usually: revenue.
  • 00:12 — Reveal: A = Costco (positive operating cash flow). B = bankrupt (revenue up, cash flow negative for 3 years). "Earnings lie. Cash doesn't."
00:15 → 01:00
Concept · 5 numbers, 3 statements
  • Story / snapshot / truth: income statement (story), balance sheet (snapshot), cash flow (truth).
  • 5 numbers: revenue growth, gross margin, operating margin, free cash flow, debt/equity. Walk through each on Costco's actual 10-K, live page-flip.
  • Pattern library: 4 mini-cases — healthy SaaS, healthy retailer, healthy bank, dying business. 90 sec each.
  • Cold-call beat: 00:55 — "Which of the 5 numbers can be made up most easily?" Acceptable: revenue (channel stuffing), operating margin (non-GAAP).
01:00 → 01:45
AI Lab · Business Health Brief
  • Solo work. Each fellow opens their pre-picked company's latest 10-K.
  • Pinned prompt starter: "Extract revenue growth, gross margin, operating margin, free cash flow, and debt/equity from this 10-K. Cite the page number for each. If a number isn't in the filing, say 'not found.'"
  • Your 3 red-pen questions (the W4 ritual): (a) Where on the page? (b) Did AI use GAAP or adjusted? (c) Is this business healthier or weaker than last year, in one sentence?
  • You rotate every 5 min asking only "Where on the page?" That is the entire week's training.
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 fellow with a clean cite, 1 who got caught uncited.
  • Standard question: "Show me the page number for your gross margin claim. Scroll to it."
  • Failure mode: fellow paraphrases AI's "approximately 30%" — make them find the exact number.
Artefact + grading1-Page Business Health Brief. Heaviest on Rigour. 9/10 = all 5 numbers cited to page, one AI rounding/paraphrase error annotated, verdict in one sentence.
Common pitfalls(1) Fellow uses 10-Q instead of 10-K — redirect. (2) AI cites a non-existent page — great teaching moment, spotlight it.
Mentor 1-liner"Where on the page? Show me the line."
W5
Sat, Sep 26 · 9 AM ET
Draft

Entrepreneurship

Learning outcome: Fellow can draft a 1-page venture with honest unit economics and name the CAC at which the business dies.

Pre-session prep · Friday

(1) Load the Airbnb 2009 deck + a real 2024 AI-wrapper deck side by side. (2) Pre-build the unit economics calculator (CAC, ASP, gross margin, payback). (3) Re-read Bill Gurley's All Markets Are Not Created Equal. (4) Brief co-mentor on capstone direction conversations — W5 is when fellows commit.

00:00 → 00:15
Hook · 2 decks
  • 00:00 — Both decks on screen, no labels. "Both raised millions. One is a real business. Which?"
  • 00:08 — Reveal Airbnb. Walk to the slide with the unit economics. "This number is the business. The other deck doesn't have this slide."
  • 00:13 — "LTV minus CAC greater than zero. That's the whole game."
00:15 → 01:00
Concept · 4 boxes + 1 equation
  • BMC compressed: only customer, value prop, channel, revenue model. Skip the other 5 boxes for a teenager.
  • Pricing as strategy: cost-plus vs. value-based vs. tiered — 90 sec each with a real example.
  • The equation: LTV − CAC > 0 and payback < 12 months. Walk through a tutoring example live.
  • The 4 deaths: no demand, no margin, no cash, no focus. Cold-call: "Name the death your venture is closest to."
01:00 → 01:45
AI Lab · 1-pager v1
  • Solo work. Each fellow drafts a 1-page venture in the shared template. Real or hypothetical.
  • Pinned prompt starter: "You are a sceptical seed investor. Read my 1-pager and give me the 5 strongest reasons this venture dies in year 1. Be specific. No platitudes."
  • Your 3 red-pen questions: (a) What's your CAC, in dollars, with the math? (b) What CAC kills the business? (c) Which of the 4 deaths is closest?
  • Capstone choice DM at 01:30: every fellow picks investment thesis / venture pitch / consumer protection artefact. You record in the cohort tracker.
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 strong venture (honest CAC), 1 fragile (no margin).
  • Standard question: "Show me the CAC. With the math. Out loud."
Artefact + gradingVenture 1-pager + unit economics sheet v1. Heaviest on Rigour + Honesty. 9/10 = stress-test passes when CAC doubles, kill-CAC named.
Common pitfalls(1) Fellow inflates LTV with fantasy retention — ask for a real comparable. (2) Fellow picks "an AI tool that does everything" — redirect to one customer, one job.
Mentor 1-liner"At what CAC does this die? Show the math."
W6
Sat, Oct 3 · 9 AM ET
Draft + Verify

AI go-to-market

Learning outcome: Fellow can answer the 5 hardest objections to their venture in under 30 seconds each, with one cited number per answer.

Pre-session prep · Friday

(1) Pre-build 5 objection persona cards (sceptical parent, competitor, VC, unhappy customer, journalist). (2) Re-read every fellow's W5 1-pager — pre-pick the hardest objection per venture. (3) Recruit one Harvard founder mentor as a guest objection-giver for the live review.

00:00 → 00:15
Hook · 2 cold emails
  • 00:00 — Two emails on screen. One AI in 4 sec, one human-rewritten in 4 min.
  • 00:06 — Reveal: human converts at 12×.
  • 00:10 — "AI drafts. Humans close. Today you become the closer."
00:15 → 01:00
Concept · GTM in one sentence
  • One sentence: "GTM is how a stranger becomes a paying customer."
  • 5 channels, 2 that work for teenagers: word-of-mouth, content. Brief the other 3 (paid, partnerships, outbound) honestly.
  • Message-market fit before product-market fit. One example: the Patagonia "Don't buy this jacket" ad.
  • The 5 hardest objections: too expensive, why now, why you, what about [competitor], how do I know it works.
01:00 → 01:45
AI Lab · objections + 3 channel messages
  • Rooms of 4 at 01:02. Each room runs the 5-objection role-play. 8 min per fellow in the hot seat.
  • Pinned prompt starter: "You are [persona]. Read this 1-pager. Give me your single strongest objection in one sentence. Then your follow-up."
  • Your 3 red-pen questions: (a) Is the answer 30 sec or less? (b) Does it cite a real number? (c) Does it concede honestly where the venture is weak?
  • Then each fellow drafts 3 channel messages (one cold email, one Instagram caption, one parent referral text).
01:45 → 02:00
Mentor review · Live objection
  • Guest mentor plays the hardest objection to 2 ventures. Unscripted.
  • Standard question: "Where did you concede? Where did you push back?"
Artefact + gradingVenture 1-pager v2 + objection appendix + 3 channel messages. Heaviest on Clarity + Honesty. 9/10 = 5 objections answered with a cited number each, one honest concession included.
Common pitfalls(1) Fellow gets defensive instead of conceding — coach: "Concede the weak point, then pivot to strength." (2) Channel messages all sound LLM-generic — force a rewrite in their own voice.
Mentor 1-liner"Concede first. Then pivot. Never argue."
W7
Sat, Oct 10 · 9 AM ET
Verify

Fraud, taxes & the real world

Learning outcome: Fellow can name the 7 Gen-Z scam archetypes, find the predatory clause in a real contract, and read a paycheck without help.

Pre-session prep · Friday

(1) Cue 60 sec of a real pig-butchering script (audio only). (2) Have a real IRS phishing voicemail recording ready. (3) Pull a real ToS (Instagram or TikTok) with the rights-grant clause highlighted. (4) Pull a real entry-level paycheck with all 4 numbers visible. (5) Safeguarding reminder: don't show real victim faces or names.

00:00 → 00:15
Hook · 3 artefacts of harm
  • 00:00 — Play the pig-butchering audio (60 sec).
  • 00:04 — Play the IRS voicemail (30 sec).
  • 00:06 — Show the ToS clause that signs away photos.
  • 00:10 — "All three target someone your age within 3 years. Today: defences."
00:15 → 01:00
Concept · 7 + 4 + 3
  • 7 scam archetypes with the one tell that breaks each: pig-butchering, fake job, fake internship, romance, crypto rug, fake refund, deepfake-of-a-friend.
  • Taxes in 30 minutes flat: W-2 vs. 1099, marginal vs. effective, the 4 paycheck numbers, why a Roth IRA at 17 wins.
  • 3 contract clauses to read first in any lease, offer, NDA, ToS, student loan.
  • Cold-call beat: 00:55 — "What's the one tell that breaks pig-butchering?" Acceptable: anyone who asks you to move money to verify it.
01:00 → 01:45
AI Lab · pick 1 of 3
  • Each fellow picks: (a) scam detector for one archetype, (b) paycheck teardown, or (c) contract redline.
  • Pinned prompt starter (a): "Give me 5 real examples of [archetype] scripts from public reporting. Cite source URLs. Identify the one phrase that always appears."
  • Pinned prompt starter (b): "Walk me line-by-line through this paycheck. Identify the 4 numbers and what each one funds."
  • Pinned prompt starter (c): "Read this ToS. Quote the 3 clauses a 16-year-old should refuse to sign. Cite the section number."
  • Your 3 red-pen questions: (a) Is every claim sourced? (b) Did AI hallucinate a statute or section? (c) Could a parent act on this in 2 minutes?
01:45 → 02:00
Mentor review · Spotlight 2
  • Spotlight 1 scam detector, 1 contract redline.
  • Standard question: "Defend one clause you flagged. Why is it predatory and not just standard?"
Artefact + gradingChoice of: Scam Detector v1, Paycheck Teardown, or Contract Redline. Heaviest on Clarity + Honesty. 9/10 = a parent can act on it without help.
Common pitfalls(1) AI invents a statute — verify against statute text live. (2) Fellow flags a standard clause as predatory — coach: "Standard isn't safe, but it isn't predatory. Show the asymmetry."
Mentor 1-liner"Could your mum act on this in 2 minutes?"
W8
Sat, Oct 17 · 9 AM ET
Defend

Capstone — build, pitch, defend

Learning outcome: Fellow pitches a real artefact in 5 minutes and survives 5 minutes of unscripted panel questions, graded live on the 4-axis rubric.

Pre-session prep · Friday

(1) Final read of every capstone draft + AI Workbook. (2) Pre-assign each fellow's slot in the spotlight schedule (6-min workshop round + defence slot). (3) Confirm outside guest panelist (CFA, founder, or faculty). (4) Print the rubric grading sheet — 4 axes, 10-point scale, 60-char comment field. (5) Test spotlight + co-host permissions.

00:00 → 01:00
Workshop · 6-min rounds
  • Each fellow gets a 6-min spotlight with one mentor. Final tightening of slides, numbers, sources.
  • Standard mentor moves: (1) cut one slide, (2) verify one number, (3) rehearse the opening line.
01:00 → 02:00
The defence · 5 + 5
  • 5 min uninterrupted pitch. Mentors do not interject. Cameras off in the audience. AI is in the room — the fellow may use it once if they choose; that counts toward the Judgment grade.
  • 5 min live questions from the panel of 3: Harvard mentor, lead analyst, outside guest.
  • The 3 standard panel questions (you rotate which one you ask): (a) What's your weakest number, and how did you stress-test it? (b) Where did AI lead you wrong, and how did you catch it? (c) If you had 8 more weeks, what's the next thing you'd change?
  • Panel grades live on the 4-axis sheet during the question round.
  • Time signals: co-host raises hand at 4:00 and 4:45 of the pitch. Strict cut at 5:00.
Grading4 axes × 10 pts each: Judgment 30% · Rigour 30% · Clarity 20% · Honesty 20%. Two mentors grade independently; outside guest tie-breaks. Pass = ≥ 28/40 with no axis below 6.
Debrief script"Kept: [strongest claim]. Cut: [weakest moment]. Push: [what to change for the next pitch in your life]."
Mentor 1-liner"You shipped it. Now own the questions."
The AI Workbook protocol

What you sign every week.
What you reject.

Every fellow keeps a single Google Doc, four columns wide. You sign each entry weekly with date and initials. No signature = artefact doesn't count.

Prompt
Raw output
Red-pen
Revision
Good · 9/10

"Explain BND in 4 sentences for a 10th grader. Cite the prospectus page for each claim."

3-paragraph answer, cites pages 12, 14, 18. One sentence says "approximately 6.5 years duration."

"Approximately" rejected. Verified actual = 6.42 years on page 14. Reworded sentence #2 in own voice.

4 sentences, exact numbers, all pages cited. AI text quoted verbatim where used.

Mid · 6/10

"Tell me about VTI."

Generic 500-word marketing-ish overview. No citations.

Notes "no citations" — but doesn't fix it.

Same paragraph, shortened. Mentor note: "Re-prompt for cited numbers. Re-submit by Monday."

Failed · 3/10

(blank)

Pasted artefact directly. No prompt, no raw output captured.

(blank)

Submitted artefact. Mentor note: "Workbook empty. Artefact doesn't count this week. Resubmit with full workbook by Wednesday or W1 grade = 0."

Feedback cadence

48 hours.
Three words.

Every fellow gets one written feedback line within 48 hours of Saturday. Format never changes: kept · cut · push. No exceptions.

The format

Kept: the one thing they did better than the cohort average. Be specific — name the line.
Cut: the one thing to remove next time. Specific. Actionable.
Push: the next stretch — one level up from where they are.

Good example · W2

"Kept: your stress table is the cleanest in the cohort — the 2008 column is exactly right. Cut: the ‘diversification reduces risk' opener — too generic. Push: rebuild the model with a fee column next week — you're ready."

Bad example

"Great work this week, keep it up!" — This is not feedback. This is a hug. If you write this, the head of pedagogy reads it.

Escalation

Miss 1 session: 30-min make-up call by Wednesday. Miss 2 in a row: flag to head of pedagogy + parent email. 3 sessions: re-entry plan for next cohort.

Mentor onboarding

Before you run a room.
The five gates.

No Harvard mentor runs a live cohort session until they've cleared all five.

01

Shadow 1 cohort session end-to-end. Camera off. Notes on. Debrief 30 min with the lead analyst after.

02

Pass AI-Workbook rubric calibration. Grade 6 sample workbooks; your scores must match the head of pedagogy within 1 point on every axis.

03

Complete safeguarding brief. Mandated-reporter training, conduct standards, the "never DM a fellow privately outside the platform" rule.

04

Run a mock 15-min Hook with the lead analyst playing the cohort. You get written feedback in the kept · cut · push format.

05

Sign mentor agreement + IP / NDA. Commit to all 8 Saturdays of the cohort and one back-up Saturday in case of illness.

Mentor recruiting · Cohort 1

If this reads like
the job you want — write us.

Harvard undergrads and graduate students only for Cohort 1. Paid. 1 cohort = 8 Saturdays + ~4 hrs prep + feedback per week. Send a 60-second video on why and a CV.

APPLY TO MENTOR Book a call