Rules · Method · Two Full Case Studies · 30 Days of Action
← → or Space to navigate · F fullscreen · 32 slides
Session 1 mapped the landscape, Session 2 put the tools in your hands — this session points all of it at the battlefield that matters most: your own career.
Entry-level jobs are shrinking, AI skills are commanding a premium, and your résumé's first screener may not be human
The AI job-search workflow: knowledge base → résumé → portfolio → interviews
Mei's road into sell-side equity research: a report pipeline
Shuai's road to Agent engineer: the portfolio is the résumé
AI amplifies the real you — it doesn't fabricate a fake one
A 30-day checklist, starting tonight
This is one of the hardest new-grad markets in decades — and the one where ordinary people hold the most leverage. Both are true at once. First, read the table you're playing at.
Sources: Stanford Digital Economy Lab · Fortune 2026-04 · Ministry of Education annual bulletins / Xinhua 2025-11 · TechTimes 2026-07
Method: the share of an occupation's tasks "already being automated by AI" in real Claude usage data — not a forecast, a measurement. Programmers rank first: exactly the reality Mei and Shuai have to face.
① Interpreters & translators ② Historians ③ Writers & authors ④ Service sales representatives ⑤ CNC tool programmers ⑥ Broadcast announcers & radio DJs ⑦ Customer service reps ⑧ Telemarketers ⑨ Political scientists ⑩ Mathematicians
The common thread: language and information processing — pure execution, fully digital, no physical presence required.
Phlebotomists, massage therapists, roofers, dishwashers, water plant operators — hands-on skill + on-site judgment: AI applicability near zero.
Editors/translators score highest of all occupations on the LLM impact index; customer service/operations postings fell over 2 percentage points in two years, sales/BD over 5 — and the bar keeps quietly rising.
Sources: Anthropic "Labor Market Impacts of AI" 2026-03-05 / Euronews 2026-03-14 · Microsoft Research arXiv:2507.07935 (checked line by line against official GitHub data) and clarification blog 2025-08-21 · PKU National School of Development × Zhaopin 2024-09
An 8× speed gap — "tailwind" isn't a metaphor here, it's data.
Sources: PwC AI Jobs Barometer 2026 official press release · WEF citing LinkedIn data 2026-01 · Xinhua / China News Service 2026-01-23
61% of employers screen résumés with AI, and 96% of recruiting professionals use AI in at least one step; AI-led first-round interviews rose from 10% to 34% in two years — more and more, the one scoring you across the keyboard is an algorithm.
81% of job-seekers already use or plan to use AI in their search; the most common use isn't "ghostwriting" but ATS compatibility checks (64%). Job-seekers who don't use AI are already the minority.
Sources: TestGorilla 2025 · Resume Now / The Interview Guys 2026 · HireVue 2026 Global AI in Hiring Report · Resume Genius 2026 · Kickresume data
AI levels the gap in "execution skills" —Execution (writing code, building spreadsheets, making decks) keeps getting cheaper with AI; judgment (what to build, is it good, why) and work you can actually show keep getting more expensive. The next five chapters teach you to put every chip on the expensive side.
and widens the gap in "judgment + portfolio + using AI well".
Remember Session 1's "advanced workflow" — knowledge base, brief, iterate, review, ship? Your job search is your first real project. Run the exact same playbook.
> Here are 100 JDs for "sell-side equity research analyst" I collected (attached). Help me: 1) extract the 15 most frequent requirements, ranked by frequency; 2) split them into "hard gates" (certificates/degrees) vs "fast-learnable skills"; 3) write the output to knowledge-base/jd-analysis.md ✓ Top 15 high-frequency requirements generated: #1 Financial modeling (mentioned in 87/100) #2 Industry research frameworks (79/100) #3 Python / data processing (61/100, up sharply from last year)…
Don't guess what the job wants — let AI count it. The high-frequency requirements are your study syllabus, and year-over-year comparison even shows where the industry is heading.
One document per target company: business lines, latest earnings highlights, executives' public remarks, campus-hiring process and past questions — your ammo depot for the night before the interview.
cron + Agent watches your target companies daily (news, filings, recruiting accounts); every change flows into the knowledge base automatically — Session 2's scheduled tasks, working for you for the first time.
Courses, projects, internships, skills, certificates — dump it all on the AI and have it re-sort everything in the recruiter's language: which are assets, which are merely experiences.
Industry × role × city × bottom lines (pay/overtime/growth). Only once written down do you notice how many "wants" fight each other — make the trade-offs before applying, not at the offer stage.
Match "what I have" against the high-frequency JD requirements and have AI produce a gap list plus a catch-up plan ranked by ROI: what can be closed in two weeks (tools, question banks), and what not to chase (gatekeeping certificates).
> Rewrite my résumé against this JD (résumé and positioning brief are in the project). Rules: 1) annotate the reason for every change; 2) quantify every result; 3) use the JD's own wording without keyword stuffing; 4) never invent any experience. ✓ 11 changes made. Example: Before: "Helped run the student club's social media account" After: "Ran the account solo for 8 months, published 32 posts, top post 12K reads, followers +40%" Reason: JD item 3 asks for "content production skills"; quantified = verifiable
A generic résumé is an average-score résumé. Have AI re-tune the emphasis and keywords for each JD — 10 minutes per version, a luxury only the AI era makes possible.
"Participated / responsible for / assisted" are dead words. For every experience, force yourself to answer: how many, how fast, up by how much — AI is great at pressing, so make it your coach.
Polishing ≠ fabricating: every number must survive a background check; never list what you didn't do. Every sentence AI writes, you must be able to defend through three levels of follow-up questions.
Note: the 87%/73% figures come from industry blog survey roundups, not primary academic reports — the magnitude is credible; treat exact values as indicative
| Company / Role | Status | Deadline | Next step |
|---|---|---|---|
| Brokerage · ER intern | Test passed | 07-20 | Interview: prep industry questions |
| Fund · Research asst. | Applied | 07-18 | Unread 3 days → find a referral |
| Bank · Fintech trainee | Rejected | — | Have AI post-mortem vs the JD |
Where you applied, current status, deadlines, next action — a spreadsheet, or let an Agent keep it; cron pings you every morning with today's deadlines.
Review conversion weekly: application→test low? Fix the résumé; test→interview low? Drill problems; interview→offer low? Practice interviews. The data tells you which stage to repair — no need for blind anxiety.
> You are a stern senior interviewer at a brokerage research institute. Dig into my résumé (in the project) section by section; after every answer, press me with two follow-ups, and at the end give: 1) per-question feedback 2) my three verbal tics 3) an improvement list Interviewer: you wrote "independently produced an industry report" — what were the data sources? How did you define the scope? If I say your valuation is 30% too high, how do you defend it?
Behavioral: STAR deep-dive follow-ups; technical: questions set from the JD; stress: have AI play the "hard-to-please" interviewer — you stay calm in the room because you've met the hard questions before.
Record live mocks or self-Q&A, transcribe, hand it to AI: verbal tics, broken logic, pacing — Session 2's video-transcription exercise, put to work here.
The night before: have AI use the company knowledge base to write 10 "questions this company is most likely to ask me" — most of your rivals never take this step.
Have AI compile the role's salary band (platform data, peer disclosures, per-capita pay in annual reports), and negotiate with a range in hand instead of being anchored to HR's first number.
Have AI play HR for three rounds: "our pay bands are standardized", "new grads have no room to negotiate" — a comeback for every script, and the cost of blowing up paid in rehearsal.
Pay, growth, platform, city, overtime — have AI build the scoring matrix, with weights you set yourself (back to the positioning brief). AI supplies structure; the human makes the choice.
Finance major, junior year. Target: sell-side equity research / investment banking. The traditional path (internship referrals + certificate stacking) keeps getting more crowded — so Mei picks a different road: let the work do the talking.
Goldman CEO Solomon: "AI can draft 95% of an S-1 in minutes — work that used to take a six-person team two weeks; the value now lies in the last 5%." Goldman's AI assistant covers all staff; JPMorgan's LLM Suite is rolled out to 250,000 employees. The junior analyst's grunt work is precisely the 95% being eaten.
24 brokerages are deploying LLMs: CICC's "Dianjing", Guotai Junan's "Junhong Lingxi", China Securities claims +40% research productivity. Campus hiring is pivoting too: Huatai's fintech track, CITIC's "Xin-Pi Program", East Money's management trainees explicitly require LLM tech + Python — while traditional IBD and research-desk openings keep shrinking.
Sources: Fortune 2025-01-17 (Solomon's own words) · American Banker · The Digital Banker · Cailian Press · 21jingji 2026-03-15 (campus-hiring requirement examples)
# Two cron jobs on Mei's server: 30 15 * * 1-5 ~/fetch-market.sh # ↑ after each close, pull quotes + sector news into the base 0 8 * * * claude -p "Read yesterday's data and write a one-page morning brief: the 3 major indexes, moves in the consumer names I track, 3 headlines and their impact" >> ~/briefs/$(date +%F).md # ↑ every day at 8:00, the brief shows up on her phone
Don't chase the whole market; pick an industry you have a feel for (Mei chose consumer). The base holds: leaders' financials, industry data sources, and the analytical frameworks from published sell-side reports.
Exercise 1 (data infrastructure) + Exercise 2 (investment analysis) bolted together are exactly this system — what was homework then is a job-hunting weapon now.
Talking shop with classmates, Mei quotes the latest numbers cold — not because her memory is great, but because her system reads the market for her every day.
The pipeline auto-pulls the target company's financials and industry data
AI runs financial analysis + comparable-company valuation on a sell-side framework
Charts + prose auto-generate a first draft; Mei owns the views and revisions
Pipeline code on GitHub, final reports on WeChat / Zhihu
One 30-page deep-dive report as the anchor, then a 2-page tracking brief every week — a continuously updated body of work proves systematic ability better than a one-off magnum opus.
Asked "can you use AI?", other candidates say "I use ChatGPT". Mei hands over a link: "Here's my pipeline — data as of this morning." In that moment, she's no longer in the same category as the rest.
❌ "Proficient with AI tools; strong interest in industry research"
✅ "Built an automated research pipeline (data ingestion → financial analysis → report generation) covering 12 consumer companies; 90 daily market briefs; deep-dive report read XX times" — every number clicks through and verifies.
How the three financial statements tie together, defending DCF assumptions under attack, the evidence behind an industry-cycle call — have AI grill you three levels deep in the target brokerage's style; then use the company knowledge base to predict questions: recent report views, the chief analyst's public calls.
CS major, junior year. Target: AI application / Agent engineer. Junior programmer jobs are shrinking, but he's aiming at the fastest-expanding slice — using work to prove he's already inside the door.
Sources: Stanford Digital Economy Lab · Final Round AI 2026 · Securities Times · Yicai · daily.dev 2026
The Hermes deployed in Session 2 plus his own accumulated skill packs (dorm-repair bot, timetable reminders) — organized into a standalone repo, with a README covering architecture and pitfalls. The tools you've used are themselves project material.
Session 1's Exercise-1 data pipeline, upgraded into a full project: scraping + cleaning + API + a one-page dashboard, with unit tests and CI — showing engineering discipline, not just features.
1–2 merged PRs to open-source projects he actually uses (OpenClaw / the Hermes ecosystem) — even doc fixes count. Recognition from a stranger maintainer beats ten thousand lines of self-praise.
# Shuai's project: the "Job-Funnel Agent" — built on his own pain point # Features: application-deadline alerts / status tracking / pre-interview # auto-summary of the company knowledge base, pushed to Telegram $ git log --oneline | head -4 e7a12f9 Add pre-interview company brief push (with tests) b3c90d1 Fix deadline timezone bug, add regression test a1f4e20 Wire up Telegram notifications 9d02c11 Initial version: funnel data model + CRUD # Ran the full Session 1 Chapter 08 workflow: # knowledge base → brief → code with tests → commit → review → ship
It solves his own real pain point (he really is job-hunting), so the requirements, iterations, and retrospectives are all real — one look and the interviewer knows it's no template project.
He opens Telegram on his phone: "These are the three deadlines it flagged for me this morning." Then the architecture diagram and the tech choices — demo + architecture + trade-offs, the interview triple strike.
The Agent is itself helping him land a job — the project and the search accelerate each other. That's topic selection at its finest.
Stuck? Don't peek at the answer — have AI give hints, not solutions ("only hint at the data-structure direction"); once solved, have it generate variant problems to test whether you understood or memorized.
Have AI play interviewer: "design a timetable-reminder system", pressing step by step on concurrency, storage, degradation — Shuai's Agent project is ready-made material, turning recital into a conversation about his own architecture.
No proxies in online assessments. Big-tech anti-cheat keeps upgrading, and one blacklist flag follows you far beyond a single company — use AI to raise your real level, not to fake one.
❌ "Familiar with ChatGPT, Claude and other AI tools"
✅ "Independently delivered a Job-Funnel Agent via an AI-pairing + test-driven workflow (Python/FastAPI, 34 commits, 85% test coverage), plus 2 merged PRs to the open-source project Hermes" — tools are worthless; what you deliver with the tools is what's worth money.
What I wrote (core logic and architecture decisions), what AI wrote (boilerplate and first-draft test cases), and how I made sure AI's code was right (review + tests + staged rollout) — those three sentences are the model answer to "can review and verify AI-generated code" in 2026 JDs.
AI adds leverage to your job search — and leverage amplifies both your equity and your losses. Three red lines to hold the floor, one mindset to hold the direction.
AI makes inventing stories easy — and makes background checks easy too. Numbers, titles, dates: one lookup and you're caught — and flagged candidates land on industry-shared blacklists.
AI proxy-taking = cheating, and anti-cheat systems (camera behavior analysis, code fingerprinting) are upgrading with AI too. You'd be competing with AI at cheating. You can't win.
"How much of this work was AI?" — the model answer is to open the workflow and walk through it. Owning your usage is a plus; concealing it is an instant veto.
AI is an amplifier:So the order cannot be reversed: first use AI to make something real (however small), then use AI to package and spread it. A portfolio is a compounding asset — Session 1 said "GitHub = the new résumé"; by now you should see it isn't the résumé's attachment. It is the résumé.
if you have work, it amplifies the work;
if all you have is blankness, it amplifies the blankness.
□ Build the knowledge base: mine 50–100 target JDs
□ Write the positioning brief (one page)
□ Gap list + catch-up plan
□ Get the cron intel station running
□ Start the killer project (report pipeline / Agent app)
□ Résumé v1 + three rounds of AI polish
□ Tidy the GitHub repos + READMEs
□ Application funnel live; apply to 3–5 a day
□ One AI mock interview per day
□ Ship v1 of the killer project
□ Post-mortem on funnel data: fix the weakest stage
□ Update the project weekly
□ Salary research + negotiation rehearsal
One Last Slide
Session 1 said AI gave ordinary people wings. Now you know: wings aren't for waiting on the wind — start the Week 1 checklist tonight.