Decide what to build, and build it all the way the AI Native way
Week 2's team project starts. In real service development, the most important thing isn't code but the order Why → What → How. Set the problem and scope from a founder's point of view, and design a 5-day sprint with gstack's Think → Plan → Build → Review → Test → Ship → Reflect sequence and AI Native development.
Thanks to AI, "how to build it" has become very fast. So if you get "what to build and why" wrong, you just build the wrong product faster.
Use the arrow buttons or keyboard arrow keys (←/→) to move. Press the button (or the N key) to open the lecture notes for each slide.
ACE AI SW Startup Training · Day 6 · Week 2 Kickoff
Build the Right Thing, then Build it Right.
People choose the right problem; people and AI together build it fast and correctly. Today we draw the full map of real service development and kick off the team project.
01 · WhyWhose problem, and what problem
02 · WhatWhat exactly to build this time
03 · HowHow to build it with the gstack sequence
04 · AI NativeHow people and AI split the work
Opening question
Why do services fail? Not "couldn't build it" but "nobody wanted it"
A common start
"It'd be cool to build this app with AI"
Write a feature list first
Draw screens first and pick a tech stack
Users are "everyone"
A start that survives
"Who struggles, when, and because of what?"
Observe how that person copes today
Start with the single narrowest user
Define success in numbers
In analyses of startup failure, "no market need" is always a top cause. AI speeds up building — and also speeds you up in the wrong direction.
Analogy · Market stalls
Not the best-made stall, but the stall where it's needed survives
Golden circle for products
Inside out: Why → What → How
HOWWHATWHY
Why · why build itWhose problem does it solve, and what problem? What evidence shows this problem is real? → PROJECT_BRIEF
What · what to buildOne key action to prove in 5 days, Must/Should/Out scope, success metrics → SITE_SPEC · USER_FLOWS
How · how to build itTech stack, data, build order, AI tools and how to verify → DATA_MODEL · IMPLEMENTATION_PLAN
What matters most
The 5 most important things in a project
1Problem firstStart from users' pain, not features
2Narrow userNot "everyone" but the one who needs it most
3Small scopeAn MVP where one key action works end to end
4Verifiable criteriaNumbers and acceptance criteria, not "make it good"
5People judgeRecord accept/change/reject of AI suggestions and why
One flow that fully solves one user's problem is a stronger demo — and a better product — than 10 flashy features.
Why · validate the problem
Validate a problem with past behavior, not opinions
Bad questions (invite compliments)
"Would you use an app like this?"
"What do you think of this idea?"
"Would you pay for it?"
Future promises and compliments are not evidence.
Good questions (invite facts)
"When did you last have this problem? How did you solve it?"
"What took the longest in that process?"
"Have you already spent money or tried something to solve it?"
Make something people want. — Y Combinator · Ask about their life, not your idea. — The Mom Test
What · define the MVP
Lock the What with 3 sentences · a scope table · success metrics
3-sentence definition① Purpose: students check in with a QR code and the teacher sees it live ② Student: search name → pick card → scan QR → checked in ③ Teacher: dashboard → live update → filter absenteesDoes the whole team picture the same product?
Must / Should / OutMust: save attendance, update status Should: late mark, filters Out: login, location check, stats reportCan you say what you're NOT building?
Success metricsStudents check in within 30 seconds Teacher sees absentees within 10 seconds 0 duplicate check-insCan you prove it with numbers in the demo?
MVP = thinnest complete slice
An MVP isn't the whole thing done roughly but one small, finished slice
Building layer by layer
Day 1: 10 empty screen shells
Day 2: all DB tables
Day 3: all APIs
Day 4: connecting... doesn't work
Day 5: no flow to demo
Slicing vertically
Slice 1: student check-in → save → shows on dashboard
(screen, data and logic all included, working)
Slice 2: duplicate/expiry exception handling
Slice 3: filters · late mark
→ a working demo every day
Every evening you need at least one "flow you can demo right now". That's the rhythm of an MVP sprint.
Analogy · From skateboard to car
MVP = a vehicle you can ride at every stage
How · the real development sequence
Real service development goes through 6 phases in order
1 · Shape the idea3 sentences on problem · users · key action, scope table, success metricsTHINK
Don't put real personal data or secret keys in docs or code
Record the accept/change/reject reason for every AI result
Takeaway
People own Why and What, people and AI do How together.
An AI Native developer isn't the fastest typist, but the person who picks the right problem, conducts the AI team and finishes with evidence.
Power to choosePick the real problem of a user who really needs it
Power to conductCall the right AI role for each phase and lock the scope
Responsibility to proveProve it's done with tests and records
1 / 24
LECTURE NOTE
Hands-on Lab
Today's Kickoff Lab Flow
See the detailed steps and done checklists for all 6 phases in the Project Day Detailed Lab Guide. Today we finish Phase 1 (idea) and start Phase 2 (planning docs).
STEP 01 · 60 min · WHY
Find the problem & interview
Brainstorm team ideas, pick one, and do 5-minute interviews with 3 users focused on past behavior.
INTERVIEW × 3
STEP 02 · 60 min · WHAT · Key
3 sentences · scope · metrics
Write the team draft first, check assumptions with /office-hours, then confirm the PROJECT_BRIEF.
PROJECT_BRIEF.md
STEP 03 · 60 min · HOW
Roles & sprint
Make a people × AI role chart, branch rules and a 5-day backlog, and install AGENTS.md and gstack in the repo.
TEAM.md · BACKLOG
STEP 04 · 40 min · SHARE
3-minute kickoff talk
Definition → scope (including Out) → metrics → roles → tomorrow's tasks. Cut scope with mentor feedback and record it.
CONVERSATION.md
gstack Command Cheat Sheet — when and why to call each
Don't accept any command's result as-is; apply it to the doc on the right with the team's judgment, then record it in CONVERSATION.md.
Command
Stage · role
When to call
Where to apply the result
/office-hours
Think · CEO office hours
Right after the 3-sentence draft. Check assumptions with 6 questions
PROJECT_BRIEF.md (problem · users · scope)
/plan-ceo-review
Plan · scope review
After the SITE_SPEC draft. Hold Scope / Reduction for 5-day projects
Scope table, SITE_SPEC feature list
/plan-eng-review
Plan · engineering review
After the 5 docs are done, and on change requests (Spec-First)
DATA_MODEL, USER_FLOWS exceptions, test plan
/plan-design-review
Plan · design review (optional)
When detailing screen design with item-by-item scores
SITE_SPEC design principles
/autoplan
Plan · run all reviews at once (optional)
When short on time. Receive only items people must decide
List of [Undecided] items
/freeze · /guard
Build · safety guard
Before building. Limit changes to one folder
Build log (allowed scope)
/design-consultation
Build · design system (optional)
When there are no color, type or spacing standards
SITE_SPEC, README
/review
Review · code review
After the core flow works, and after building a change
Build log, fix commits
/qa-only → /investigate → /qa
Test · QA
Report only → investigate root cause → fix · re-check · regression
TEST_PLAN.md, BUG_LOG.md
/document-release
Ship · update docs
Before the final talk. Match the docs to the final code
README.md, the 5 planning docs
/retro
Reflect · retro
At the end of the sprint
RETROSPECTIVE.md, talk retro
Kickoff Template Kit — copy and use right away
Copy each block with its Copy button and save it as a file in your team repo. Change the [bracketed] parts for your team.
WHY · WHAT
PROJECT_BRIEF.md
Today's key output. Every Phase 2 doc uses this file as input.
# [Service name] — PROJECT_BRIEF
## 1. 3-sentence definition
1. Purpose: a service that lets [who] do [what].
2. [User A]: [start] → [action] → [result]
3. [User B]: [start] → [action] → [result]
## 2. Evidence of the problem (interviews · observation)
| # | Person | Recent situation | How they solve it now | Pain point |
|---|---|---|---|---|
| I-1 | | | | |
| I-2 | | | | |
| I-3 | | | | |
## 3. Key user (the one who needs it most)
- Who:
- When · where:
- How they cope now (Status quo):
## 4. Scope
- Must (no demo without it):
- Should (if time allows):
- Out (not building this time + why):
## 5. Success metrics (in numbers)
- [User A] completes [action] within [N seconds/N steps]
- [Errors · duplicates] [N] or fewer
## 6. Risks and open questions
- [Assumption]:
- [Question]:
## 7. Applying AI review (/office-hours)
- Accepted:
- Changed:
- Rejected + why:
WHY
5-Minute User Interview Notes
The Mom Test principle: ask about the person's past behavior, not your idea.
# INTERVIEW I-[number]
- Person: [role, situation] · Date: [ ] · Interviewer: [ ]
## Questions (don't explain the idea)
1. When did you last face [problem situation]?
2. How did you solve it then? Can you walk me through it in order?
3. What took the longest or was most annoying in that process?
4. Have you tried any tools or methods to solve it?
5. What would change if that problem were solved?
## Notes (separate fact / interpretation)
- Fact (exactly as heard):
- Interpretation (what we think):
- Surprises (Observation & surprise):
## Verdict
- This person has the problem [a lot / a little / hardly]
- Evidence:
HOW · AI NATIVE
TEAM.md — people × AI role chart
Name one final approver for each role.
# TEAM — roles and teamwork rules
| Role | Person | Decisions owned | Paired AI |
|---|---|---|---|
| PO · planning | [name] | Why/What, scope approval, change requests | /office-hours, /plan-ceo-review |
| Design · UX | [name] | User flows, screen standards | /design-consultation, /plan-design-review |
| Dev lead | [name] | Structure · data, code merges | Codex · Antigravity, /plan-eng-review, /review |
| QA · release | [name] | Test verdicts, demo prep | /qa-only, /qa, /document-release |
## Teamwork rules
- No direct commits to main · feature/[REQ-ID]-[summary] branches
- One REQ = one plan = one commit
- Integrate at least once a day, demo every day at 17:00
- The approver checks the diff before merging AI output
- Record AI use in CONVERSATION.md
## Daily scrum (10 min)
- REQs finished yesterday / REQs to finish today / blockers
HOW
SPRINT_BACKLOG.md — 5-day plan
Arrange it so you have one more "demo-ready flow" every evening.
# SPRINT BACKLOG — Week 2
| Day | Phase | Goal (evening demo) | REQ / output | gstack | Owner |
|---|---|---|---|---|---|
| 6 Mon | Think · Plan | Brief confirmed, kickoff talk | PROJECT_BRIEF, TEAM | /office-hours | |
| 7 Tue | Plan | 5 docs + mid-term talk | SITE_SPEC, USER_FLOWS, DATA_MODEL, IMPLEMENTATION_PLAN | /autoplan | |
| 8 Wed | Build · Review | 1 core flow works | FR-001 ~ FR-00_ | /freeze, /review | |
| 9 Thu | Test · Iterate | POC talk, fix defects | TEST_PLAN, BUG_LOG, CHANGE_REQUEST | /qa-only, /qa | |
| 10 Fri | Ship · Reflect | Final presentation | README, RETROSPECTIVE | /document-release, /retro | |
## Definition of done (all REQs)
- Acceptance criteria (Given-When-Then) reproduce in the browser
- 0 console errors, mobile and PC layouts work
- Merged after review, logged in CONVERSATION.md
Log
CONVERSATION.md — decision history
Not a chat copy, but question → key suggestion → team decision → reason → evidence.
Used from POC feedback (Day 9). Write it before the code.
# CR-001 — [change request in one sentence]
- Requested by · date:
- Reason (user value):
- Agreed real meaning: [e.g. real time = 5-second polling]
- Done criteria (Given-When-Then):
## Impact analysis
| Area | Impact | Docs/files to change |
|---|---|---|
| Screens | | |
| Flows | | |
| Data | | |
| Tests | | |
| Schedule | | |
## Decision
- This sprint / next version backlog
- Approver:
- Order: fix docs → /plan-eng-review → approval → code → /review → /qa
prompt
gstack · AI Prompts by Phase
Fill the [brackets] and paste into Codex or Antigravity chat.
[THINK] /office-hours
We want to build a service that solves [problem] for [users] in 5 days.
Ask us questions based on the 3-sentence definition and interview notes below,
and point out the riskiest assumption and the narrowest starting scope.
[paste 3-sentence definition + interview summary]
[PLAN] /plan-ceo-review
Review SITE_SPEC.md in Hold Scope mode. Suggest features that can't be demoed
in 5 days as candidates to move to Out, and tell us if any core value is missing.
[PLAN] /plan-eng-review
Review the 5 planning docs, list contradictions between docs, hidden assumptions and missing exception flows in a table,
and create a test plan for each requirement ID (FR-xxx).
[BUILD] Build request (Codex · Antigravity)
Read the 5 planning docs first and propose only the build plan for FR-[ ].
Show the files to create or change and why, and the acceptance criteria to check; build after approval.
[REVIEW] /review
Review the website/ code against SITE_SPEC's requirement IDs and acceptance criteria.
Look first for features not in the plan, and unhandled save failures and double clicks.
[TEST] /qa-only
Run the TEST_PLAN.md scenarios in the browser; don't fix any code,
and leave reproduction steps and screenshots for each defect.
[REFLECT] /retro
Using CONVERSATION.md and the work history, summarize what went well, what was hard,
cases where we rejected or changed AI suggestions, and 1 improvement for the next version.
Don't add achievements without evidence.
SHARE
3-Minute Kickoff Talk Script
For today's wrap-up talk. If the scope is big, the mentor cuts it down right there.
# Kickoff talk (3 min)
0:00 Problem (40s)
- "[Users] face [pain] in [situation]."
- One line of interview evidence: "I-2: ..."
0:40 What (60s)
- 3-sentence definition
- 3 Must / 3 Out (+ why they were cut)
1:40 Success metrics (30s)
- "[action] within [N seconds]", "0 [errors]"
2:10 How · team (30s)
- Role chart (people × AI), sprint goals for Days 7–10
2:40 Questions · requests (20s)
- 1 decision you want the mentor to confirm
Click the buttons above to go to the Day 1–10 learning pages.
Learning Workspace
Startup · gstack · AI Native Videos & Day 6 AI Helper
Pick a reference video by topic on the left, and ask the AI chat on the right about Why·What·How, the gstack sequence and running an AI Native team.
Related YouTube VideosDay 6 recommended materials
Why · What · startup thinking
1. Start with why — how great leaders inspire action (Simon Sinek)The Golden Circle · where Why → What → How beginsEN
2. Everything We Teach at YC in 10 MinutesY Combinator's key advice in 10 minutesEN
3. How to Get and Test Startup Ideas — Michael SeibelFinding problems and testing ideasEN
4. How to do customer interviews? — Rob Fitzpatrick (The Mom Test)Interview questions that get facts, not complimentsEN
5. What a Lean Startup MVP (minimum viable product) meansThe MVP concept, in KoreanKR
MVP · scope
6. Michael Seibel — How to Plan an MVPA YC partner on setting MVP scopeEN
7. How to Build An MVP (Y Combinator)Build small and launch fastEN
8. How to succeed at startup MVP developmentStartup MVP development, part 1KR
9. Key strategies for MVP developmentStartup MVP development, part 2KR
gstack dev sequence
10. Analyzing gstack, the Claude Code plugin made by YC's CEOgstack structure and skills explained in KoreanKR
11. How to install gstack, the open-source hit in KoreaInstall and basic useKR
12. YC CEO Open-Sourced His Entire Claude Code Setup — GStackIntro to the full gstack workflowEN
13. I Tried Garry Tan’s GStack, Claude Code Skills That Run a StartupReal use from /office-hours to /qaEN
14. gstack Explained: Is It Right for You?When to use it, and when not toEN
AI Native development
15. Andrej Karpathy: Software Is Changing (Again)Software 3.0 · when natural language becomes the programEN
16. What Is an AI-Native Engineering Team?Roles and ways of working in an AI Native teamEN
17. Maximizing AI-native development skills with GitHubAI-native development in practice (Korean)KR
18. What INTENT.MD is and the AI-Native way of developing [KR/EN subs]AI Native development centered on intent docsKR
19. Antigravity in practice & SDD AI developmentDoc (spec)-driven AI development (Day 4 review)KR
Tech reference (earlier Day 6 videos)
20. Learn JavaScript - Full Course for Beginners (freeCodeCamp)Build basics · JavaScriptEN
21. WEB3 - Intro to Ajax (Opentutorials)Async communication conceptsKR
22. WEB3 - Using the Ajax fetch API (Opentutorials)Sending and receiving data with fetchKR
23. JavaScript - Intro to jQuery Ajax (Opentutorials)jQuery Ajax referenceKR
Videos are embedded public YouTube videos. If playback is restricted by the uploader, open them directly on YouTube. gstack and AI tools update often, so commands and screens in videos may differ from the current version.
Day 6 · Daily Feedback
Today's Course Satisfaction Survey
After Day 6 class, please answer a short survey. Your answers help us improve the next class and the course.