탈락하는 발표
화면은 예쁘지만 “누가 실제로 쓸지는 아직 찾아보는 중입니다”
목요일 · 발표까지 하루001 / 95
목요일 — 발표까지 단 하루
오늘 하루 동안 내일 5분 발표에 필요한 모든 것을 구축합니다. 완성된 앱이 아니라 논리적 서사 + 증거 + 보여줄 1개의 화면을 만듭니다.
문제 검증→스코프 축소→증거 수집→시연 동선→발표 구상
목요일 · 발표까지 하루002 / 95
오늘의 타임테이블
| 시간 | 구분 | 마칠 때 손에 쥐어야 할 결과물 |
|---|---|---|
| 10:05–11:10 | A — Interrogate the Problem | 한 문장 가설 |
| 11:10–11:50 | B — Cut Scope + Send Questions | 핵심 기능 1개 확정 · 질문 2건 전송 |
| 11:50–12:50 | Lunch (답변이 돌아오는 시간) | — |
| 12:50–13:30 | C — One Demo Path | 동작하는 화면 또는 목업 |
| 13:50–14:30 | D — Design the Talk + Pair Rehearsal | 5슬롯 발표 대본 |
The gstack intro and install were finished on Monday — 10:00–10:05 is only a show-of-hands check. Grey rows: intro · lecture · rehearsal · wrap-up
목요일 · 발표까지 하루003 / 95
내일 평가받는 항목
“무엇을 만들었는가”가 아니라 “무엇을 새로 발견했는가”입니다.
데모 시연은 주인공이 아니라 증거 자료입니다.
합격하는 발표
화면은 투박하지만 “2명에게 물어보니 내 가설이 틀렸음을 확인했습니다”
목요일 · 발표까지 하루004 / 95
솔직히 말씀드리면 — 이는 본 교육과정 범위를 넘어섭니다
- 비즈니스 모델 이론은 본 교육과정 평가 항목이 아닙니다. BM 이론으로 점수를 매기지 않습니다.
- 대신 평가하는 것은 문제를 얼마나 날카롭게 정의했는가, 현실의 사용자를 직접 만났는가, 그리고 한계를 솔직하게 밝혔는가입니다.
- 오늘 배우는 것은 스타트업 창업론이 아니라 AI를 활용해 판단의 질을 높이는 방법입니다. 이것이 바로 우리 수업의 핵심 목표입니다.
목요일 · 발표까지 하루005 / 95
오늘 완성할 4가지
① 한 문장 가설
누가 · 언제 · 왜 · 어떤 손실
② 2 pieces of evidence
실제 대상자에게 질문하고 인터뷰한 기록
③ 시연 동선
3번의 클릭으로 충분함
④ 5슬롯 발표 대본
내일 5분 발표의 뼈대가 되는 구조
These four are the entire day. The urge to build a fifth thing is today’s biggest risk.
목요일 · 발표까지 하루006 / 95
손들기 점검 — 3가지 질문
- Q1 이번 학기 500줄 이상의 코드를 작성해본 적이 있나요?
- Q2 낯선 타인에게 자신의 아이디어를 설명해본 적이 있나요?
- Q3 그 사람이 “나도 그게 필요해”라고 말하는 것에서 그치지 않고, 이미 스스로 대안을 찾고 있는 것을 확인해본 적이 있나요?
손이 내려가는 속도가 오늘 수업의 핵심 메시지입니다.
목요일 · 발표까지 하루007 / 95
2일 스프린트의 핵심 명제
구축 비용이 0으로 수렴할수록, 모든 가치는 “무엇을 만들지 선택하는 능력”으로 이동합니다.
목요일 · 발표까지 하루008 / 95
오늘의 도구 — 설치하거나, 안 되면 프롬프트를 붙여넣으세요
You finish the install on your own, Monday or Tuesday — no commands to memorise: paste one paragraph and Codex runs it for you. If it does not work, you go with the workbook prompts. Whichever you use makes zero difference to today’s grade.
목요일 · 발표까지 하루009 / 95
gstack이란 무엇인가 — 90초 요약
“It turns Claude Code into a virtual engineering team — a CEO who rethinks the product, an eng manager who locks architecture, a designer who catches AI slop, a reviewer who finds production bugs, a QA lead who opens a real browser…”README.md, github.com/garrytan/gstack
Y Combinator CEO인 Garry Tan이 공개한 MIT 라이선스 오픈소스입니다. AI에게 새로운 기능을 추가하는 것이 아니라, AI에게 역할(Role)을 부여하는 규칙 모음입니다.
The quote says “Claude Code” because that is the author’s default environment. Ours is the OpenAI Codex app (desktop). gstack supports Codex, but recognition inside the app is not guaranteed — which is why the prompt track is our default.
목요일 · 발표까지 하루010 / 95
단순 질문 vs 역할 부여 비교
단순 질문 방식
“예약 앱 만들어줘”
→ 300 lines, instantly
→ “no, that’s not what I meant…”
→ another 300 lines
→ binned two hours later
역할 부여 방식
“당신은 냉철한 투자 분석가입니다. 칭찬은 생략하고 제 문제 정의부터 엄격하게 비판하세요.”
→ 6 questions
→ “That is hearsay. It is not your own experience”
→ the hypothesis changes
All gstack does is pre-write the sentence on the right for 50 different situations. There is no magic. Which is exactly why we can copy that sentence and use it ourselves.
목요일 · 발표까지 하루011 / 95
대부분은 마크다운(Markdown) 프롬프트입니다
“gstack gives Claude Code a persistent browser and a set of opinionated workflow skills. The browser is the hard part — everything else is Markdown.”ARCHITECTURE.md
즉, 하나의 명령어 = 하나의 마크다운 파일입니다. .agents/skills/ 디렉토리에 마크다운으로 존재하므로, 복사-붙여넣기 방식과 완전히 동일하게 동작합니다.
목요일 · 발표까지 하루012 / 95
직접 열어보기 — 5분이면 충분합니다
⌥github.com/garrytan/gstackMIT 라이선스 · 전체 소스 공개 · 클릭시 새 탭에서 열기Open exactly two things.
ETHOS.md— what this tool believes. A 10-minute read- Any folder →
SKILL.md— the “wait, that’s all?” moment is today’s core. It is just a Markdown prompt
목요일 · 발표까지 하루013 / 95
제작자의 시연 영상 보기
▶
How to Make Claude Code Your AI Engineering Team — Y Combinator · Garry Tan
Click to play it right here (from the 2:45 mark)youtube.com/watch?v=wkv2ifxPpF8&t=165s
Click to play it right here (from the 2:45 mark)youtube.com/watch?v=wkv2ifxPpF8&t=165s
If you are offline, open the URL directly — youtube.com/watch?v=wkv2ifxPpF8
목요일 · 발표까지 하루014 / 95
영상에서 이 3가지에 집중하세요
- The questions come first — the scene that starts with questions, not code. You will live through the same thing in 구분 A today
- The roles switch — the same AI is a CEO, then a reviewer. Because a human assigns the role
- The human rejects — the scene where he takes a proposal and says “no”. That is everything you will learn today
Your eyes will drift to the flashy automation, but what we are stealing today is ① and ③.
목요일 · 발표까지 하루015 / 95
설치 — 이 문단을 그대로 복사해 붙여넣으세요
Windows / WSL2 Instructions:
First, check the current Windows environment. If WSL2 is installed, run the installation procedure below inside WSL2. If WSL2 is not available, verify whether this project officially supports PowerShell. Adapt any macOS/Linux-specific commands for Windows as needed, and if any step fails, identify the root cause and resolve it. Do not overwrite or delete any existing **gstack** files in the user's home directory without explicit permission.
# Copy everything inside the box below, paste it into the Codex chat, and press Enter.
Run these in order. If anything fails, find the cause and fix it.
1) If bun is missing, install it: curl -fsSL https://bun.sh/install | bash
2) git clone --single-branch --depth 1 \
https://github.com/garrytan/gstack.git ~/gstack
3) cd ~/gstack && ./setup --host codex
4) If it asks about the naming scheme, choose option 1 (it does not affect Codex)
# When it finishes, restart the Codex app and type $office in the chat
Office Hours YC Office Hours — two modes. (gstack) ← if this appears, it worked
문단을 단축하거나 수정하지 말고 그대로 붙여넣으세요. 설치가 안 되면 워크북 프롬프트를 이용하세요. 결과와 평가 점수는 동일합니다.
목요일 · 발표까지 하루016 / 95
윈도우(Windows 환경) 환경 — 3가지 사전 준비
| 필요 항목 | Mac 환경 | Windows 환경 |
|---|---|---|
| Shell | Terminal — already there | Install Git Bash separately git-scm.com/download/win |
| Node.js | Not needed | Needed nodejs.org |
| bun | curl -fsSL https://bun.sh/install | bash | In PowerShellpowershell -c "irm bun.sh/install.ps1|iex" |
| Running the install | Terminal, as-is | In a Git Bash window (not PowerShell) |
- Shortcut — if Node.js is already there, one line works on both:
npm install -g bun - bun needs Windows 환경 10 1809 or later
Honest advice: starting from zero on Windows 환경 does not finish in 10 minutes. In class, go with Track B (paste the prompt) and install after school. Tool usage carries zero points in today’s grading.
목요일 · 발표까지 하루017 / 95
슬래시(/)가 아닌 $ 기호로 호출합니다
As written in the gstack docs
/office-hours · /review
— that is Claude Code notation
In the Codex app, actually
Type $ in the chat and a list appears
Pick from the list — no names to memorise
- Type just
$offand pick Office Hours. Same method for$qa·$review - The names in the list are short (Office Hours, Qa, Review). The folder names carry a
gstack-prefix, but you can ignore that when picking - Even without calling the name, Codex sometimes picks the skill up on its own when the content matches (implicit invocation)
- Watch for lookalikes —
Gstack Openclaw Office Hoursis a variant for a different agent. What we use is plainOffice Hours
For readability, this guide also writes /office-hours. In the app: $off → pick from the list. If nothing appears, paste the workbook prompt — the result is the same.
목요일 · 발표까지 하루018 / 95
Codex 앱 환경에서 달라지는 6가지
Self-study — we skip this in class today
| Feature | Claude Code | Codex (us) |
|---|---|---|
| Multiple-choice questions | AskUserQuestion buttons | None. Comes out as a “Decision Brief” in prose |
/careful /freeze | Real blocking via hooks | No hooks. Advisory wording only |
| Second opinion | /codex | gstack-claude (needs the Claude CLI) |
| Parallel expert review | Runs | Removed |
| Invoking a skill | /name | $name or auto-detection |
| In the app (GUI) | — | Personal skills may not appear in the list (open issue) |
The second row matters most today — on Codex, the guardrails block nothing. We come back to this later.
목요일 · 발표까지 하루019 / 95
ETHOS — 오늘 꼭 기억할 단 하나의 원칙
“AI 모델은 제안하고, 결정은 사용자가 합니다. 이 원칙은 다른 모든 규칙에 우선합니다.”ETHOS.md
“Two AI models agreeing on a change is a strong signal. It is not a mandate.”
The top principle of gstack’s three. It is today’s tiebreaker for every conflict — “the AI told me to” is not a reason.
목요일 · 발표까지 하루020 / 95
AI는 여러분의 CTO가 아닙니다
AI가 하는 일
Questions, alternatives, drafts, reviews, tireless checking
여러분이 하는 일
Choose the problem, decide the scope, approve and reject, secure the evidence, take the responsibility
AI를 CTO라고 부르고 싶다면, 여러분은 CEO입니다. 그리고 CEO의 핵심 역할은 거절(Reject)하는 것입니다.
목요일 · 발표까지 하루021 / 95
오늘 사용할 명령어 — 6개면 충분합니다
| 구분 | 기본 방식 — 워크북 프롬프트 붙여넣기 | 선택 사항 — 설치된 경우 |
|---|---|---|
| A 문제 검증 | P1 six-question prompt | $office-hours |
| B 스코프 축소 | P3 scope-cutting prompt | $plan-ceo-review |
| C Mockup | P5 self-contained HTML prompt | $design-html |
| C Code check | P6 two-pass checklist | $review |
| C Screen check | P7 — make your neighbour click through it | $qa |
| C Cross-review | Swap code with your neighbour | Not today — your neighbour’s eyes are faster |
The left column alone completes today. The right column is a bonus. The repo has 50+ commands; today we use six.
/office-hours /plan-ceo-review /design-html /review /qa gstack-claude
목요일 · 발표까지 하루022 / 95
오전에 절대 하지 말아야 할 일
금지 사항
새로운 기능 개발하기
디자인 다듬기
기술 스택 고민하기
gstack 명령어 구경하기
권장 사항
문제 정의 문장 다듬기
사용자에게 연락하기
기능 과감히 덜어내기
The morning’s two hours succeed or fail on how much you deleted.
목요일 · 발표까지 하루023 / 95
이 도구는 단순 개발 툴이 아닙니다
Building the app now takes a single afternoon. The hard part is choosing what to build — and you do that with questions, not code. The three lines on the left are this entire morning.
목요일 · 발표까지 하루024 / 95
한눈에 보는 역할 분담
목요일 · 발표까지 하루025 / 95
스킬 카탈로그 — 단계별 활용 가이드
The next 8 slides are reference. In class we only walk through ① and ②. ③–⑥ and the “5 worth knowing” are material you open when you need it — not exam material.
gstack’s skills are laid out in the order you build a product. Don’t memorise them — just take away “ah, this stage has one of these”.
| Stage | Flagship skill | One line | Today |
|---|---|---|---|
| Think | office-hours | Pressure-test the idea before code | ✅ 구분 A |
| Plan | plan-ceo-review | Decide whether scope grows or shrinks | ✅ 구분 B |
| Build | design-html | One screen, made real | ✅ 구분 C |
| Review | review | Interrogate my code like a stranger’s | ✅ 구분 C |
| Test | qa | Click through it like a user | ✅ 구분 C |
| Ship·Reflect | ship retro | Deploy · retrospect | Not today |
목요일 · 발표까지 하루026 / 95
① office-hours — pressure-test the idea
When
Before a single line of code. The moment you think “should I build this?”
Why
Thinking alone, you interpret everything in your own favour. You need a cross-examiner
How — it fires 6 questions one at a time, and re-asks until your answer gets specific.
you ▸ You are a cold-blooded early-stage investment analyst. No compliments.
Idea: helping students find someone to eat with during free periods
Skip product design — attack the problem definition first.
ai ▸ Q1. What evidence do you have that someone would be genuinely stuck without this?
you ▸ My friends say it’s annoying.
ai ▸ That is hearsay. Has anyone already spent time or money on it?
Where it helps: it tells you in 10 minutes that you mistook a solution for a problem. Alone, that takes two days.
목요일 · 발표까지 하루027 / 95
② plan-ceo-review — decide the scope
When
After deciding what to build, before building it
Why
Left alone, AI always expands. A disaster for anyone with a deadline
How — name one of the 4 modes. Skip it and it drifts toward expansion. EXPANSION (grow) · HOLD SCOPE (freeze) · SCOPE REDUCTION (cut — us, today)
you ▸ Mode: SCOPE REDUCTION
Constraints: 40 minutes today, working alone
Success metric: the single scene I show in tomorrow’s talk
Cut everything that does not serve it. Keep a list of what you cut.
Where it helps: the cut list becomes your “What we got wrong” material, verbatim.
목요일 · 발표까지 하루028 / 95
③ design-html — one screen, made real
When
When there is nothing to show yet. The night before the talk
Why
Words alone are not believed. One clickable screen beats ten slides
How — demand HTML that opens as a single file. Name the 4 states (empty/loading/results/error) or it will not look real.
you ▸ Build a one-page self-contained HTML file.
No external CDNs or images. It must open as a single file.
Required states: empty / loading / results / error
Fill it with realistic fake data — real names, real sentences. No lorem ipsum.
Where it helps: a presentable demo in 40 minutes, no backend. Most of you take this path today.
목요일 · 발표까지 하루029 / 95
④ review — my code, through a stranger’s eyes
Self-study — we skip this in class today
When
Right before committing code, or the pre-talk check
Why
You cannot see your own code. And you especially cannot see code an AI wrote for you
- How — two passes: critical findings first, quality later
- Good at catching: places where model-generated values get saved or sent with no validation
- What humans miss: a new state value whose consumer lives in another file — invisible in the diff
you ▸ For every finding, quote the exact code as evidence
and rate your confidence out of 10. Do not fix anything — list only.
Careful: left alone it fixes things on its own (commits are qa’s job) — hence “do not fix”.
목요일 · 발표까지 하루030 / 95
⑤ qa — click through it like a user
Self-study — we skip this in class today
When
Right before showing the demo to anyone else
Why
Unit tests check “is my thinking right”; QA checks “what the other person experiences”. They are not substitutes
How — normally it opens a real browser, clicks around, and keeps screenshots. Today, a human is faster.
Today’s recommended method — hand your laptop to your neighbour, state only the goal with zero explanation, and let them click for 3 minutes. Watch where they stall.
Better than the tool — because a person stalls exactly where an explanation is needed.
Careful: never run it on a deployed live service — it is a real browser with real cookies and sessions, so payments and deletions really fire.
목요일 · 발표까지 하루031 / 95
⑥ safety tools — and their limits
Self-study — we skip this in class today
| Skill | What it does | On Codex |
|---|---|---|
careful | Asks before rm -rf, DROP TABLE, force-push | No hooks, so no enforcement. Only advisory wording appears |
freeze | Locks edits to one folder | |
guard | Both at once |
“The guardrails are accident prevention, not access control.”docs/skills.md
Even on Claude this was “accident prevention”, never “security”. On Codex, even that does not engage. Today the only real defence line is one thing: commit often.
목요일 · 발표까지 하루032 / 95
Not used today — but 5 worth knowing
Self-study — we skip this in class today
| Skill | When to use it | Why not today |
|---|---|---|
spec | Turn “something like this, roughly” into an executable document | 5 stages — does not fit in 40 minutes |
investigate | Bugs. Never fix without a cause | Today there is no time even to create bugs |
cso | Security audit (OWASP + threat model) | A mockup has nothing to audit |
ship | Tests→review→PR. Stops on failure | We do not go as far as deployment |
retro | Retrospect a week in numbers | Two days is too short |
These 5 pay off on a personal project after the semester. Especially the iron rule of investigate — “no fix without a cause” — a discipline you can use with no tools at all.
목요일 · 발표까지 하루033 / 95
구분 A — Interrogate the Problem (65 min)
- 30 min solo — interrogate your own idea with
$office-hoursor the workbook P1 prompt - 20 min in pairs — attack each other’s hypotheses (10 min each)
- 15 min whole class — only the 4 people whose hypothesis changed present + instructor walk-around wrap-up
Target output: one-sentence hypothesis + 1 riskiest assumption. Get those out and 구분 A has succeeded.
/office-hours
목요일 · 발표까지 하루034 / 95
/office-hours — what it does
A YC partner office-hours session, turned into a command. It pressure-tests the idea before a single line of code gets written.
“Push on each one until the answer is specific, evidence-based, and uncomfortable. Comfort means the founder hasn’t gone deep enough.”office-hours/SKILL.md
Two Codex-app cautions: ① No choice buttons — the questions arrive as prose in the body, and you answer in your next message. ② Installed: $office-hours. Not installed: paste workbook P1 as-is. The result is the same.
/office-hours
목요일 · 발표까지 하루035 / 95
What you actually type
# Installed: $office-hours / otherwise paste the P1 prompt and start
you ▸ $office-hours
Problem: students in my department can’t find anyone to eat with
during free periods, so they settle for snacks alone at the campus shop.
Skip product design — attack the problem definition first. No compliments.
ai ▸ Q1. What is the strongest evidence that someone would be
genuinely stuck if this went away?
(“it would be nice to have” is not demand)
you ▸ My friends said it’s annoying.
ai ▸ That is hearsay. It is not your own experience.
Has anyone already spent time or money on this problem?
/office-hours
목요일 · 발표까지 하루036 / 95
Q1 — is the demand real
“What’s the strongest evidence you have that someone actually wants this — not ‘is interested,’ but would be genuinely upset if it disappeared tomorrow?”
Not demand
“My friends said it would be nice”
“80% said they need it in our survey”
“They say the market is huge”
Demand
Someone is already spending money or time
You can name — by name — the person who would be stuck without it
목요일 · 발표까지 하루037 / 95
Q2 — how are they coping right now
“What are your users doing right now to solve this problem — even badly? What does that workaround cost them?”
Red flag: “Nothing exists. That’s why the opportunity is so big.”
If truly nobody is doing anything, that problem does not hurt enough. One person hand-building their own spreadsheet is a stronger signal than 100 survey responses.
목요일 · 발표까지 하루038 / 95
Q3 — name that person
“Name the actual human who needs this most. What’s their title? What gets them promoted? What gets them fired? What keeps them up at night?”
Red flag: category-level answers — “university students”, “students living off campus”, “small business owners”.
“These are filters, not people. You can’t email a category.”office-hours/SKILL.md
Minimum bar: resolution like “a 300-level business student with a 3-hour Tue/Thu gap between lectures and a 90-minute commute”.
목요일 · 발표까지 하루039 / 95
Q4 — the narrowest wedge
“What’s the smallest possible version of this that someone would pay real money for — this week, not after you build the platform?”
Extra pressure: what if it created value with the user doing nothing at all? No login, no setup, no integrations.
The answer here becomes the only screen you build in 구분 C. Write it down word for word.
목요일 · 발표까지 하루040 / 95
Q5·Q6 — observation, and 3 years out
Q5 Observation
“Have you watched someone use it without helping? What surprised you?”
“I ran a survey” is not an answer — Surveys lie. Demos are theater.
Q6 3 years out
“As the world changes over 3 years, does this become more necessary or less?”
“The market grows 20% a year” — Growth rate is not a vision.
목요일 · 발표까지 하루041 / 95
A bad start vs a good start
Bad
“Build me a lunch-buddy matching app for free periods. In React.”
A solution mistaken for a problem. The AI writes 300 lines immediately, and you throw them away two hours later.
Good
“/office-hours
Problem: 300-level business students with a 3-hour Tue/Thu gap can’t find anyone to eat with, so they settle for snacks alone at the campus shop.
Skip product design — attack the problem definition first.”
/office-hours
목요일 · 발표까지 하루042 / 95
Default track — paste exactly this
# Paste this into the Codex app chat as-is. No install needed.
You are a cold-blooded early-stage investment analyst. No compliments.
My idea: [one sentence]
Ask me the following 6 questions, one at a time. Do not move to the
next question until I have answered. If my answer is vague, ask again.
1) What is the strongest evidence that someone would be genuinely stuck without this?
2) How is that person coping with the problem today, and what does that cost them?
3) Describe that person specifically — not as a category.
4) What is the smallest version someone would pay for this week?
5) Have you watched someone use it without helping? What surprised you?
6) Three years from now, is this more necessary or less?
At the end, score my hypothesis out of 100 and justify the score.
목요일 · 발표까지 하루043 / 95
No idea yet — six starters
① Cafeteria · free periods
Eating alone, cafeteria queues, menu info
② Campus secondhand
On-campus resale deals, scams and no-shows
③ Club money
Cash and transfers mixed up — can anyone trust the club dues ledger?
④ Tutoring · study groups
Matching tutoring gigs, scheduling, drop-outs
⑤ Cooking on a budget
Splitting groceries, expiry dates, group buying
⑥ Gym · fitness
Waiting for gym equipment, workout partners, tracking
If you are picking one, pick it here, within 10 minutes. And only pick something where you can reach the actual people involved this afternoon.
목요일 · 발표까지 하루044 / 95
The one-sentence hypothesis — template
[WHO] , during [WHEN · WHAT SITUATION] ,
because [WHAT BLOCKS THEM] suffers [WHAT LOSS] .
# Example
A 300-level business student who commutes , during the 3-hour Tue/Thu gap ,
because they can’t find anyone to eat with on the spot ,
settles for snacks at the campus shop and loses afternoon focus .
If a product name appears in this sentence, it is wrong. The solution does not enter yet.
목요일 · 발표까지 하루045 / 95
One riskiest assumption
If the hypothesis collapses, where exactly does it collapse?
This afternoon’s 2 real-world contacts must be an attempt to break this one assumption. Ask random questions and you only burn time.
목요일 · 발표까지 하루046 / 95
Pair cross-examination — the 20-minute rules
- 10 min each, then swap. The speaker reads the one-sentence hypothesis, then stays silent
- The listener asks questions only. Proposing ideas is banned
- Three required questions — “How would you know if this is wrong?” / “Has anyone already spent money or time on this?” / “Who will you ask besides people you know?”
- Final minute: the listener repeats the hypothesis back in one sentence. If they can’t, that hypothesis is still blurry
목요일 · 발표까지 하루047 / 95
구분 A — to pass this block
A-1 You have a one-sentence hypothesis (no product name)
A-2 One riskiest assumption is written down
A-3 2 named people to contact this afternoon are chosen
A-4 Your pair could repeat your hypothesis back
If any of the four is missing, fill it before lunch. Do not move on to 구분 B.
목요일 · 발표까지 하루048 / 95
구분 B — Cut Scope + Send Questions (40 min)
- 20 min the P3 prompt (or
$plan-ceo-review) in SCOPE REDUCTION — cut down to 1 feature - 10 min write the 3 questions you will ask this afternoon
- 10 min actually send them — WhatsApp, DM, phone call. Now.
Sending is part of 구분 B. “I’ll send it later” means it never gets sent. Lunch becomes the hour the replies come in.
/plan-ceo-review
목요일 · 발표까지 하루049 / 95
/plan-ceo-review — one mode of four, only
| Mode | What it does | Today |
|---|---|---|
| SCOPE EXPANSION | “What gets 10x better for 2x the effort?” | ❌ |
| SELECTIVE EXPANSION | Cherry-pick expansions to keep | ❌ |
| HOLD SCOPE | Freeze the scope, edge cases only | △ |
| SCOPE REDUCTION | Cut to the minimum version. “Be ruthless.” | ✅ this one only |
“Once the user selects a mode, COMMIT to it. Do not silently drift.”plan-ceo-review/SKILL.md
/plan-ceo-review
목요일 · 발표까지 하루050 / 95
The scope-cutting prompt (shared by Track A/B)
/plan-ceo-review ← Track B: drop this line, keep the rest
Mode: SCOPE REDUCTION
Constraints: 40 minutes this afternoon, working alone, dev experience [high/mid/low]
Hypothesis: [one sentence]
Success metric: [the single scene I will show in tomorrow’s 5-minute talk]
Cut everything that does not serve this metric.
Keep the cut items in a list, each with the reason it was cut.
If what remains is more than 2 screens, cut again.
The cut list becomes the material for tomorrow’s slot 4. Do not delete it.
/plan-ceo-review
목요일 · 발표까지 하루051 / 95
Today’s hard limits
1core feature
2screens, max
3clicks make the demo
40minutes to build it
If it can’t be built in 40 minutes, it is not what you build today. Cutting scope is skill, not surrender.
목요일 · 발표까지 하루052 / 95
The ladder of evidence — today’s target is ③
④ and ⑤ are impossible in one day. Today’s target: two pieces of ③. And tomorrow, saying “we only got as far as ③” scores higher than pretending you reached ⑤.
목요일 · 발표까지 하루053 / 95
The 3 questions to ask — and the banned ones
Ask
“When did you last run into that situation?”
“What did you do then?”
“Have you ever spent money or time because of it?”
금지 사항
“Would you use an app like this?”
Everyone says “yes”. Zero information
“What do you think of this feature?”
Ask for a verdict and you get politeness back
The principle: ask about past behaviour, not future intent.
목요일 · 발표까지 하루054 / 95
Friend bias — write it down, don’t hide it
The 2 people you contact today will almost certainly be people you know. That is not the problem. Pretending otherwise is.
Points off
“We validated with 2 users” (both are roommates)
Points on
“Both are friends of mine, so there is bias. Here is how I discounted their answers accordingly”
목요일 · 발표까지 하루055 / 95
The evidence log — use exactly this format
| # | Who | Friend? | Actual behaviour (past tense) | Where it differs from my assumption |
|---|---|---|---|---|
| 1 | A, 300-level, same department | Yes | Snacked alone 3 times last week. Asked the class WhatsApp group, got no reply, gave up | The problem may be response rate, not matching |
| 2 | B, junior from my club | Yes | Eats alone on purpose — finds it easier | Falsifying signal — not everyone is bothered |
If the last column is empty, that interview was never heard. You only went to get your own opinion confirmed.
목요일 · 발표까지 하루056 / 95
What lunch is for
11:50–12:50 is not just for eating — it is when the replies come in.
- Sending finished inside 구분 B. The replies arrive over lunch
- No reply? Send to 1 more person at the start of the afternoon
- Reply arrives? Fill the last column of the evidence log immediately
This design is the only channel through which reality intervenes in these two days.
목요일 · 발표까지 하루057 / 95
구분 B — to pass this block
B-1 Cut down to 1 feature · 2 screens at most
B-2 The cut list survives, with reasons
B-3 All 3 questions ask about past behaviour
B-4 Actually sent (verified by screenshot)
목요일 · 발표까지 하루058 / 95
구분 C — One Demo Path (40 min)
You already built something
Check it with P6 → confirm the path with P7 (neighbour clicks) → keep only the one path you will show and hide the rest
Nothing yet
P5: one static screen + fake data. 40 minutes is enough
Both branches share one goal — the 3 clicks you show for 90 seconds tomorrow.
/review /qa /design-html
목요일 · 발표까지 하루059 / 95
Nothing built yet — the 40-minute mockup prompt
/design-html ← Track B: drop this line, keep the rest
Build a one-page self-contained HTML file.
No external CDNs, fonts, or images. It must open as a single file.
Screen: [e.g. — find someone to eat with during a free period]
Required states, all four: empty / loading / results / error
Fill it with fake data that looks real — real names, real sentences.
No lorem ipsum.
Check that nothing breaks at 375px mobile width and report back.
The lorem ipsum ban is not about taste — it is about layout lies. Real sentences run longer than lorem and break layouts differently.
/design-html
목요일 · 발표까지 하루060 / 95
AI slop — what erodes trust tomorrow
Slop signals
Overdone purple-teal gradients
Meaningless emoji icons
Every card with the same shadow, same rounding
Leftover placeholder text (“Lorem ipsum”, “John Doe”)
Instead
Set the layout with real sentences
Build the empty and error states first
Colour only where it carries meaning
Slop is not an aesthetics problem — it is a trust problem. The moment it reads as “an AI made this”, nobody hears your content.
목요일 · 발표까지 하루061 / 95
Already built — what /review catches
- SQL · data safety string interpolation, writes that bypass validation
- Race conditions duplicate submissions, unsafe HTML rendering
- LLM output trust boundaries model-generated values flowing into the DB, mail, fetch — with no validation
- Shell injection executing string-interpolated commands
- Constant completeness tracing a new value to consumers outside the diff
In student projects the real catches are 3 and 5. Number 5 is a defect humans structurally cannot catch — it never appears in the diff.
/review
목요일 · 발표까지 하루062 / 95
Auto-fix — a commit you haven’t read is not your code
/reviewauto-fixes some findings. The list includes N+1 query fixes and added LLM-output validation — both can change behaviour/qafixes and then commits- The brakes are rule-based counters —
qastops above 20% risk · hard cap 50 findings. Yet the basis for those coefficients appears nowhere in the repo
The rule: read every auto-fix commit line by line with git show. Tomorrow, saying “the AI fixed this part and I did not check it” beats presenting without knowing.
/review /qa
목요일 · 발표까지 하루063 / 95
Codex has no hooks
Claude Code
/careful·/freeze are PreToolUse hooks. The tool call itself gets confirmed or blocked
Codex (us)
The hooks: frontmatter gets stripped. What remains is one paragraph of “Safety Advisory” prose
“The guardrails are accident prevention, not access control.” … “it’s accident prevention, not a security sandbox.”docs/skills.md
Even on Claude, the guardrails were never security controls. On Codex, even that weak device has no force. The only thing standing between you and rm -rf today is your own hands.
목요일 · 발표까지 하루064 / 95
/qa — it opens a real browser
“Repro is everything. Every issue needs at least one screenshot. No exceptions.”qa/SKILL.md
- Unit tests check “is my thinking right”;
/qachecks “what the user experiences” - It demands a clean working tree — commit first, then run it
- Never run it on a deployed live service. It is a real browser with real cookies and sessions
Short on time? Track B is enough — hand your laptop to your neighbour and let them click with zero explanation. 3 minutes surfaces most of the problems.
/qa
목요일 · 발표까지 하루065 / 95
“Where this demo is lying, right now”
Write down everything fake on your current screen. Tomorrow you present this list, as-is.
# The fake list (example)
· Recommendation algorithm: none — it is just sorted by most recent
· The 3 users are dummies I created
· Login does not work — it is only a button
· Notifications appear on screen only. Nothing actually gets sent
Hide it and get caught in Q&A, and that talk is over. Say it first, and it becomes honesty points.
목요일 · 발표까지 하루066 / 95
Lock the one demo path
Start screen→Click 1→Click 2→Result
- You must be able to show these 4 steps in 90 seconds, without speaking
- If it needs explanation midway, it is not a path — it is an excuse
- Practise 3 times and delete any click that fails
- Check laptop brightness, resolution and notifications-off — today
목요일 · 발표까지 하루067 / 95
구분 C — to pass this block
C-1 A click path that finishes inside 90 seconds
C-2 The fake/unbuilt list is written down
C-3 Opened once on someone else’s laptop or phone
C-4 (Optional) a 3-minute backup recording
C-4 is optional — but if the network dies tomorrow, it saves the talk.
목요일 · 발표까지 하루068 / 95
Tomorrow’s 5 minutes — 5 slots
Slot 4 decides the character of these two days. Empty, it is an investor pitch; filled, it is a validation report.
목요일 · 발표까지 하루069 / 95
Five things to delete
- Tech-stack explanation → delete (“we used all six gstack commands” is also a delete)
- Market size with no basis → delete
- The full screen tour → shrink to the one path
- A roadmap of unbuilt features → delete
- No quotes → add 1 direct quote, mandatory
Take these five out of 5 minutes and 3 minutes remain. Those 3 minutes are the real content.
목요일 · 발표까지 하루070 / 95
Do not say “we validated it”
Asking 2 people cannot verify a hypothesis. All you can do is attempt to falsify it.
Wrong wording
“We validated user needs”
“Market fit has been proven”
Accurate wording
“1 of the 2 people gave us a falsifying signal”
“We tried to break the hypothesis; under these conditions it held”
목요일 · 발표까지 하루071 / 95
Pair rehearsal — the two-run rule
- Run 1 time it, nothing else. Over 5 minutes? Shrink slot 3 (the demo)
- Run 2 the listener throws 2 Q&A questions — one is mandatory: “How would you know if this is wrong?”
- Stuck on an answer? Write it into slot 5 (next validation) on the spot
A question that stumps you is not something to hide — it is material for slot 5. Know it in advance and Q&A stops being scary.
목요일 · 발표까지 하루072 / 95
Tomorrow’s operations — know this in advance
| 시간 | What |
|---|---|
| 09:30–09:50 | Opening · grading criteria · peer feedback cards handed out |
| 09:50–11:50 | Session A — 8 people (15 min each: talk 7 + Q&A 5 + changeover 3) |
| 11:50–12:50 | Lunch |
| 12:50–12:53 | Morning-patterns briefing (3 min) |
| 12:53–14:23 | Session B — 6 people |
| 14:23–14:40 | Peer feedback tally · best comments |
| 14:40–15:00 | Wrap-up — how to distrust your tools |
The order is drawn by lot tomorrow morning. Everyone arrives by 09:30 with laptop, adapter and backup file ready.
목요일 · 발표까지 하루073 / 95
Tonight’s checklist · the last word
- 한 문장 가설 (final)
- 2 evidence log entries — including the last column (where it differs from my assumption)
- 90-second demo path + the fake list
- 5슬롯 발표 대본 + 2 rehearsal runs
- Laptop · adapter · backup file (or the recording)
- Notifications off — a popup mid-talk costs you trust
Presenting “this turned out to be wrong” tomorrow is not a penalty. Finding that out in one day is an achievement.
The penalty is for inventing results that do not exist.
금요일 · DEMO DAY074 / 95
금요일 — 데모데이
What we grade today is “what you found out”, not “what you built”. The demo is not the star of the show — it is evidence.
14students
7 mintalk
5 minQ&A
0points for polish
금요일 · DEMO DAY075 / 95
How today runs
| 시간 | What |
|---|---|
| 09:30–09:50 | Opening · draw the speaking order · hand out peer feedback cards |
| 09:50–11:50 | Session A — 8 people (15 min each) |
| 11:50–12:50 | Lunch |
| 12:50–12:53 | Morning-patterns briefing (3 min) |
| 12:53–14:23 | Session B — 6 people |
| 14:23–14:40 | Tally the peer feedback cards · read out the best comments |
| 14:40–15:00 | Close — how to doubt your tools |
15 min per person = 7 talk + 5 Q&A + 3 transition. When the timer rings at 7 minutes, you stop — even mid-sentence. No exceptions.
금요일 · DEMO DAY076 / 95
평가 루브릭 — 100점 만점
| Criterion | Points | What we look for |
|---|---|---|
| Problem clarity | 20 | Is it specific — who, when, what loss? Naming a category costs you points |
| Evidence of real contact | 25 | Did you actually ask people · did you capture past behaviour · did you disclose how many were friends |
| Hypothesis updates | 20 | What changed because of what you heard. If nothing changed, you owe us a reason |
| Demo fit | 20 | Does the scene you show help confirm the hypothesis (not how polished it is) |
| Honesty | 15 | Did you name the fake, the unbuilt, the limits first. Bonus for self-reporting |
“Polish” and “business-model theory” carry zero points. Because that is not what this course is about.
금요일 · DEMO DAY077 / 95
디자인 완성도 점수가 0점인 이유
If we graded polish
Whoever spent one more day coding yesterday wins → the lesson of these two days is wiped out
If we grade evidence
Whoever asked a real person yesterday wins → we are grading the skill that survives the AI era
As the cost of building falls toward zero, all the value moves into “the value of choosing what to build”.
금요일 · DEMO DAY078 / 95
동료 피드백 카드 — 3가지 소평가
| Box | What to write |
|---|---|
| ① The strongest line | The most convincing sentence in the talk — copy it down word for word |
| ② The weakest link | The spot where “problem → solution” or “evidence → conclusion” made a leap |
| ③ One question you want to ask | Even if you never got to ask it in Q&A, write it here |
No scores. When students score each other, relationships contaminate the grading. We collect sentences only.
금요일 · DEMO DAY079 / 95
청중의 역할 — 3가지 질문 유형
- The falsifying question “How would you know if that was wrong?”
- The evidence question “Of the people you asked, had anyone already spent money or time on this?”
- The honesty question “Which part of this demo is fake?”
Of the 5 Q&A minutes, at least 2 questions come from you. If the instructor asks everything, you become spectators.
금요일 · DEMO DAY080 / 95
발표자 — 발표 30초 전 점검사항
- Kill notifications (Slack · WhatsApp · email)
- Prefer plugging in directly over screen sharing — no lag
- Have your backup file or recording path already open
- Say your first sentence out loud once — it opens up your voice
Your first sentence is not a greeting — it is the problem sentence. Do not burn 15 seconds on “Good morning, my name is…”.
금요일 · DEMO DAY081 / 95
발표 5슬롯 구조 — 다시 확인하기
- 0:00–0:45 The problem + 1 direct quote from an interview
- 0:45–1:30 How the hypothesis changed — at first → now → what changed it
- 1:30–3:00 The demo — 3 clicks, one path
- 3:00–4:00 What we got wrong
- 4:00–5:00 The next test
The slot is 7 minutes, so you have 2 minutes of slack. That slack is for when the demo stutters — not an invitation to explain more.
금요일 · DEMO DAY082 / 95
세션 A 시작
8 people from here. 15 minutes each.
- No questions during a talk. Write them on your card
- The next speaker plugs their laptop in ahead of time
- Keep the 7-minute timer visible on screen
금요일 · DEMO DAY083 / 95
Right after lunch — morning-patterns briefing (3 min)
We stop for just 3 minutes here and name what kept repeating across the morning’s 8 talks.
Repeated weaknesses
(Instructor fills this in live — e.g. friends-only interviews, demo tours, saying “we validated it”)
Repeated strengths
(e.g. leading with the disconfirming case, self-reporting the fake list)
The whole point of these 3 minutes: the afternoon 6 do not repeat the morning’s mistakes.
금요일 · DEMO DAY084 / 95
세션 B 시작
6 people left. You are allowed to apply the morning’s feedback immediately.
- Fixing your script is not cheating — it is learning
- But if you changed something, say so: “I watched the morning talks and changed this part” — that earns bonus points
금요일 · DEMO DAY085 / 95
What the good talks had in common
- A specific person shows up in the first 30 seconds — “a 300-level Business Admin student who commutes”
- The quote comes out verbatim — “He said, ‘I’d just go to the campus kiosk’”
- The demo is one path. Not a tour
- They say what they could not build first
- “What to check next” is measurable
금요일 · DEMO DAY086 / 95
가장 흔한 5가지 실수
- The category user — “Students find it inconvenient”
- The intent interview — “They said they would use it” (future intent is not data)
- The demo tour — clicking through all 5 screens until time runs out
- “We validated it” — you cannot validate anything with 2 people
- Hiding the fake — get caught in Q&A and that talk is over
금요일 · DEMO DAY087 / 95
검증(Verify) vs 반증(Falsify)
Impossible
Asking 2 people in order to verify
“The need has been proven”
Possible
Trying to break your hypothesis
“1 of the 2 people gave me a disconfirming signal, so I changed the hypothesis to this”
This distinction is the longest-lasting thing you learn today — because it is training in judgement, not in entrepreneurship.
금요일 · DEMO DAY088 / 95
이제 도구 자체를 의심해볼 시간입니다
For two days you used Codex and gstack. The last 20 minutes go to how to doubt these tools.
① Framing with numbers
What “810×” really is
② Auto-fix
Commits that change behaviour
③ Silence ≠ safety
The blind spots of checking tools
The tools will be different in 2 years. The method of doubting will not.
금요일 · DEMO DAY089 / 95
① Don’t quote “810×” as-is
The gstack repository offers a comparison: “one person’s output grew 810×”. And then the same repository takes that number apart itself.
| Where you poke the assumption | Multiplier |
|---|---|
| The README’s original claim | ~810× |
| Raise the 2013 baseline from 14 lines → 50 lines | 228× |
| Correct for 2× AI verbosity (the author’s own figure) | 408× |
| Push the correction to 100× | 8× |
“시간 to first user is the metric that matters, not LOC.”docs/ON_THE_LOC_CONTROVERSY.md
금요일 · DEMO DAY090 / 95
② Auto-fix and outsourced judgement
/review·/qafix the code and commit it. The list includes items that change behaviour- The brake is a rule-based counter, and the docs give no basis for its coefficients
- gstack’s top ETHOS principle is “AI models recommend. Users decide.” — yet the automation features, by definition, replace that very decision
The lesson: that a guardrail exists and how strong its basis is are two different questions. A number attached is not a number verified.
금요일 · DEMO DAY091 / 95
③ Silence is not safety
- The security check
/csocarries 22 hard exclusion rules — DoS, rate limiting, and missing logging are structurally never reported - The sentence the tool stamps onto every report, without exception: “This tool is not a substitute for a professional security audit… produce false negatives.”
- The same tool behaves differently per environment — Codex has no hooks, so
/carefulblocked nothing at all
A pass from an automated checker means “not found yet”. It does not mean “not there”.
금요일 · DEMO DAY092 / 95
Bonus — the tool doesn’t match its own docs either
| Source | Count |
|---|---|
| What the README says | “23 specialists + 8 power tools” = 31 |
| Commands the install guide lists | 35 |
| Directories that actually contain a SKILL.md | 54 |
A tool that promises to stop documentation drift is drifting in its own README. We hit it ourselves — the README lists --host cursor, but the install script rejects that value as an error.
The code is the truth, not the docs.
금요일 · DEMO DAY093 / 95
So the real metrics are
0points for polish
2real-world contacts, minimum
1next test
1honest failure report
시간 to your first user is the metric. Not lines of code.
Today’s rubric and this sentence are saying the same thing.
금요일 · DEMO DAY094 / 95
Three things you take with you — 3 minutes to write
- 1 thing to keep A command or habit that actually paid off over these two days
- 1 thing to drop Something you only did for form — write its name down
- 1 thing to count next time What will you measure in your next project
The gap between people who use the tool as-is and people who bend it into their own workflow opens up 6 months from now. Every file in ~/.codex/skills/ is plain markdown — open it and edit it.
금요일 · DEMO DAY095 / 95
최종 명제
Build a beautiful app and say “we’re not sure yet who will use it” — that talk fails.
Build an ugly mockup and say “I asked 2 people and 1 of them showed me I was wrong” — that talk succeeds.
Codex and gstack are just the operating system that runs this loop. People make the decisions, all the way to the end.
AI models recommend. Users decide.