SESSION 06 / 06 — FINAL
Remote setup · Six-session review · Next steps
Jee Jun Young
Business major → Journalism · Overseas sales → Software development
iOS · Android · Frontend About 5½ years (sixth year)
Developed iOS/macOS apps for an enterprise collaboration tool
Since 2026 Solo web and server development and commercialization through vibe coding
PART 1 — Three things today
REMOTE — WHY
Commuting, cafés, bed — sooner or later you will want Claude working at those times too.
01 — AMPHETAMINE
Free on the Mac App Store.
A MacBook normally sleeps when closed, and then cannot be reached remotely.
A small utility that keeps it awake even with the lid closed.
No need to memorize. Just know “this exists.”
At home, ask Claude: “I want remote access to my MacBook. Tell me whether I need something like Amphetamine and guide the setup.”
02 — REMOTE ACCESS
No need to memorize. Just know there are three routes.
Ask Claude at home: “Which remote-access option suits my setup: SSH, VNC or a cloud IDE? Assess it.” Like the technology-choice self-check in session 5.
Terminal → move to the working folder → enter claude remote
→ The session appears as a chat in the Mac app. Continue giving instructions there.
Benefit: easy to continue the same session on mobile; terminal, Mac app and phone form one workflow.
⚠ I have not tried Cursor remote. Paid remote tools were slow, so I use them only in emergencies.
03 — MOBILE
REMOTE × CLAUDE FEATURES
Shared across iPad, Mac and iPhone. Like KakaoTalk.
My use: simple questions, writing fiction with Projects, and researching personal-website examples.
In your code folder (or the parent folder containing web/server projects), set up remote access → it appears in the Claude iPad/iPhone app.
Queue up small instructions and check progress during drives or tea breaks.
Not a direct connection to Cowork itself: go to the folder Cowork is working in, set up remote access, then it appears in the iPad/iPhone app for small instructions.
My use: presentation drafts and technical reviews of lecture content and slides.
PROMISE FROM SESSION 01
— Session 1
“When you wake up, step out of the shower, or return from coffee, has something been made?
A report, a spreadsheet summary, something organized? Probably not.”
So we spent 6 weeks on vibe coding to build “a system with something finished when you return from coffee.”
There is another key to building that automated system.
Harness engineering and agents.
HARNESS ENGINEERING · AGENT
— Mitchell Hashimoto · 2026.2.5 · My AI Adoption Journey
“Whenever an agent makes a mistake, design a mechanism that prevents the same mistake from happening again.”
The creator of Terraform and Vagrant. Officially adopted by OpenAI within 2 weeks; Anthropic and Martin Fowler followed with publications.
The core equation
Agent = Model + Harness
If the model (Claude/GPT) is the brain, the harness is the environment that makes it work properly: ① constraints against going off task, ② limiting it to the few tools it should use, ③ specifying what it must and must not do.
AGENTS.md, tool connections, automated tests and error-recovery loops: a vast set.
My own work in progress
I am refining harness and agent concepts in my Grimoire project. The central effect: instead of the developer fixing Claude’s mistakes directly, Claude checks and critiques its own code and documents.
Developers already apply parts of this, knowingly or not. How many have built it systematically enough for Hashimoto or Martin Fowler to say “You are really doing it properly” is unclear. Knowing these two terms and their roadmaps is what enables a work environment where humans do little and AI works precisely.
PLAIN — TWO SENTENCES
Recognize the limits
Even with dozens of documents and a well-established canonical source, a moment comes when Claude turns stupid.
Remember the developer-chat screenshot from session 3: “Claude has gone crazy. Where did the old one go?”
Harnesses and agents, in plain language
Instead of me telling Claude “Do it again”,
Claude checks, corrects and retries its own work.
A loop where AI checks and corrects its output without a human stepping in every time. That is the essence of Hashimoto’s “mechanism that prevents the same mistake twice.”
That is why, in Grimoire,
this is exactly what I am refining.
REMOTE — IN PRACTICE
Three examples: me, the bookkeeper from my last company, and a friend.
CASE 01 — Me
My route
Before dawn, leave on the motorcycle with iPad and iPhone
→ Reach Onam Lake Park in Namyangju and stare at the lake
→ Review Claude’s report on my phone and plan the next instructions
→ Ride alone to the Seoul Museum of Craft Art in Anguk
→ Catch my breath under the night sky → give Claude more work.
The workflow in one line
Terminal → working folder → claude remote → a chat appears in Mac/mobile apps → continue instructions there.
That is Claude remote.
Leave with only your iPhone/iPad and keep the MacBook awake with Amphetamine. It all works.
CASE 02 — The bookkeeper from my last company
What they are doing
Remotely restoring a classic fighting game from 20 years ago.
Uses all tokens on Claude’s top plan. Company work stays in a separate company Claude account.
Checks the Mac mini and Windows PC only after work; during weekdays, monitors remote tasks in progress.
Now
Moved from business support at a spatial-design firm to head of an AX team.
AX means AI Transformation: AI strategy departments Korean large and midsize companies began creating in 2024–2026, following digital transformation (DX).
That is Claude remote.
iPhone/iPad + Amphetamine + a remote session. It all works.
CASE 03 — A friend, and wrap-up
A friend, in week 3 of vibe coding
He cannot distinguish Claude-generated HTML from a web developer’s web project. Two hours of tutoring did not clarify canonical sources and references. He even raised technically impossible objections about iOS push notifications.
But he is automating his own job. He is building a webpage that calculates and organizes overseas-goods quotes for an exporter, work previously done by several field staff and team leads measuring dimensions and accounting for damage.
Where he gives instructions: cafés, bars, anywhere — iPhone/iPad + Amphetamine + remote.
These are very basic examples of “having Claude work while I rest or am elsewhere”.
You can start immediately when you get home.
Add more developed methods such as harness engineering, agents and Mac minis, without omissions, bugs or errors, and…
→ I work only 30 minutes, then just review and approve agent reports while the business runs.
That is how the startup CEO in the KBS documentary from session 4 and real experts charging hundreds of thousands of won for courses direct work.
You are standing at that starting point too.
PART 2
I am not trying to scare you.
Just giving you a direction.
6 AREAS
You do not need to know everything. When needed, ask Claude directly.
ASK CLAUDE
THE POINT
you move ahead of
everyone who has not even started vibe coding.
After the break: reviewing all 6 sessions.
PART 3
6 SESSIONS — RECAP
And the 10 principles are on the website.
GUIDE — Things to try at home ①
① CRUD practice: try different things
CRUD = Create, Read, Update, Delete, the backbone of nearly every web service.
Try building small things: a guestbook, simple reservations, a site for your designs or photos.
② Vercel/Supabase self-check
“Can my entire service be built with Vercel and Supabase?”
If not: “What should I use for the server: Azure, AWS or GCP?”
③ Compare hosting and servers (good for CLI work)
“Among Cloudflare and Korean hosting providers, which can you work with best through CLI?”
“How do I connect my webpage to proper hosting and a server? Separate what I must do myself from what you can do through CLI.”
⚠ Key point when working on servers, infrastructure and hosting
If you click through AWS/Azure page by page, asking “What is this?” for every term, you will burn out.
Ask Claude about unfamiliar words, and tell it to perform the actual configuration and work through CLI. That is the way.
GUIDE — Things to try at home ②
① Ask about the vocabulary
“What does native development mean?”
“What does web hybrid mean?”
“Why use Flutter? How does it differ from web hybrid and native?”
② App-release process
“How do I create iOS/Android accounts and release an app?”
Have Claude guide each step: Apple Developer/Google Play Console signup, review and app submission.
⚠ Note: initial iOS/Android release setup is difficult
This is an area where outsourcing/turnkey vendors give online-shop owners guides longer than 10 pages.
If the owner cannot understand the guide, the vendor does it for them for a service fee. The process is that complex; Claude’s step-by-step guidance has real value.
6 WEEKS — WHAT WE COVERED
Definition of vibe coding (Karpathy’s February 2025 tweet) · My session 1 message: “There is no one-click coding”
Vercel/Supabase deployment · Why documents matter, through the restaurant analogy · Four principles: act as the client, document, do not take AI at face value, delegate
Canonical sources and pyramids (CLAUDE.md, PLANNING.md) · Context Rot: context decay and attention loss · Separation of Concerns (SoC) · Keep each file small
F12 and 6 defenses (RLS, .env, NEXT_PUBLIC_, BFF, source maps, .gitignore) · Public/secret key self-check · account.md / SECURITY.md · AI’s average cases vs. hackers’ edge cases
MVVM, FSD, Clean Architecture, SOLID (vending machine, department store, onion, sushi kitchen) · CLI as Claude’s hands and feet · Four technology-choice scenarios · “People with the words direct AI” (Wittgenstein, 1922)
Remote setup (SSH, VNC, cloud IDE) · `claude remote` workflow · Harness engineering and agents (Hashimoto, February 2026: Agent = Model + Harness)
That is what you explored in 6 weeks.
It is okay to forget. If a term is blurry, just ask Claude again.
PART 4
At the start of session 1,
I said this.
WHERE YOU ARE NOW
You are ahead of
everyone who has not even started vibe coding.
Q&A · INDIVIDUAL
Questions are also welcome in the KakaoTalk open chat.
You can stay for an individual consultation.
Thank you for your work over these 6 weeks.