I direct AI to build AI. No coding background, one company, lots of mistakes, all public.
Seoul, Republic of Korea
Joined September 2025
- Tweets1.7K
- Following274
- Followers999
- Likes398
Pinned Tweet
9 months ago I had zero coding experience.
Today an AI I built runs daily work at my company: reports, customer-service drafts, shipping, document search.
I built it all by directing Claude. Here's the whole story, mistakes included π§΅
8
6
226
An error can look exactly like an empty inbox.
In a night test, my company's AI read 13 of 121 broken responses from online stores as "no new questions."
The fix cut that to 3 and put a red warning above the waiting list.
Building this in public. More of the mess soon.
1
1
15
Mr. ZON π°π· retweeted
μμ¦ μμ€λ‘
"1μ² λ¬λ¬ μ΄μ μ£ΌκΈμ μ΄λ»κ² νλ©΄ λ°μ μ μλμ?" λλ "λͺ μκ°μ ν¬μ νλ©΄ κ·Έλ§νΌ λ°μ μ μλμ?" κ°μ λ΅κΈμ λ§μ΄ μ£ΌμλλΌκ³ μ.
κ·Έλμ μ°Έκ³ μ©μΌλ‘ μ§§κ² κΈμ λ¨κΉλλ€.
μ£ΌκΈμ λ§μ΄ λ°κΈ° μμνλ©΄μ λλμ보λ,
ν°μ κΉλ €μλ μ± μ€ λνλ―Όμ κ°μ₯ ν¬κ² ν°λ¨λ € 쀬λ μ νλΈλ₯Ό λ³Ό μκ°λ κ΄μ¬λ μ μ μμ΄μ§λλΌκ³ μ.
κ·Έλ κ² ν°μ μ μ μ νμ λ¬Έμλ₯Ό νμΈνλ μ©λλ‘λ§ μ¬μ©νκ³ μμκ³ , μμ€λ νμ λ°±κ·ΈλΌμ΄λ μνμμ μ΄μ΄μ ν¬μ€ν
νκ³ λ΅κΈ λ¬κ³ μμ΅λλ€.
κ·Έλ¦¬κ³ λμ΄ λ»κ·Όνκ³ ν° νλ©΄μ 보λκ² κ΅¬ν μ 릴λλ μμ΄μ. κ·Έλ§νΌ ν° λ³΄λ μκ°μ κ±°μ μμ€μ μ¬μΈνλ μ λλ‘ νλ€λ³΄λ©΄ μ£ΌκΈμ μ¬λΌκ°λλ€.. (μ½μ§λ μλ€μ.)
λΉμ₯μ νλ‘μλ μ κ³ λ
ΈμΆλ μλκ³ λ΄ ν¬μ€ν
μ μ¬λλ€μ λ΅κΈλ μ λ€κ³ λ무 μμν΄ λ§μκ³ , κΎΈμ€ν νλ€ λ³΄λ©΄ λΉμ λ°ν λ μ΄ μ€μ€κ²λλ€. νμ΄ν
! π₯
United Airlines just blocked AI assistants from booking its flights.
They can still show you fares. A person has to buy the ticket.
Part of the reason is check-in. A human has to confirm the bags hold no explosives.
Some steps still need someone who can make a promise.
17
Mr. ZON π°π· retweeted
λΆμκ³Ό κ±±μ μ΄ λ¨Έλ¦Ώμμ κ½ μ±μΈ λ λΉμ₯ μ’
μ΄λ₯Ό κΊΌλ΄λΌ.
λ¨Έλ¦Ώμμ λ λλ μκ°μ μ€μ²΄λ³΄λ€ 10λ°°λ λ 컀 보μΈλ€.
μ’
μ΄μ λ΄κ° μ§κΈ ν΅μ ν μ μλ κ²κ³Ό
ν΅μ ν μ μλ κ²μ λ°μΌλ‘ λλ μ μ΄λΌ.
ν΅μ ν μ μλ 건 κ³Όκ°ν μ°’μ΄λ²λ¦¬κ³
ν΅μ ν μ μλ λ± ν κ°μ§λ§ μ§κΈ ν΄λΌ.
π€ Made with AI
When an AI says it reviewed a document, did it look at the pictures?
Mine said it checked everything. It had read the file as text only.
The 9 images inside held three requests that weren't in the text. Staff had drawn them.
Now it opens every image first.
12
A date is only right if you know whose clock wrote it.
My sales sheet should date parcels by the day they shipped. It read US Pacific time, and 11:55 PM there is already the next afternoon in Korea.
Every US parcel came out a day early. Now it uses the shipper's own date.
19
κ΅³λͺ¨λμ
λλ€
λͺ¨λλ€ νκΈ°μ°¬ ν루 λμκΈΈ.
μ’μμΌμ λ ν¬κ²,
λμμΌμ λ μκ² λκ»΄μ§μκΈΈ.
All is Well.
A report that has to say something every day needs a way to say nothing.
The first thing I built was a Telegram bot that sent me a market briefing every morning.
Version 8 got a new rule. If nothing changed, say "No new signal."
Still building it. More of the mess soon.
39
Telling an AI "this codebase is yours" changes how it works.
In Anthropic's tests on Claude Fable 5.2, that sentence got it to clean up repeated code about half the time. Without it, under 10%.
It also made more changes nobody asked for.
Do you let your AI own the code?
You can lose data just because two people use different words for the same thing.
On the customer screen, the answer was saved.
On the staff screen, it showed blank. That screen didn't know the word.
One save and it was gone.
Staff testing our claims form caught it.
1
23
Something that says "failed" might have already worked.
A 30 min timer on my AI only killed the part watching an update.
The update finished minutes later with nobody looking.
31
i had zero coding experience when i started building my company's AI
so i didn't tell it how to build things
i handed it the project and asked it to approach it like a top expert in that field would
if you can't give the steps, borrow the expert's thinking
1
57
my AI could search 441,552 pieces of our company docs
it only matched exact words
ask with different words than the doc and it missed
matching words isn't understanding the question
took till late may to make it search by meaning
50
building my company's AI with Claude Code, no coding background.
our daily inventory report kept going out late or on stale numbers.
cause: the data pull wrote "done" before checking if it worked, so a failed try still said done.
now "done" means it really worked.
6
4
92
my AI was dead for 26 hours and nothing told me.
the check only asked "is it switched on?"
it was. it just wasn't answering anyone.
"on" and "working" are not the same thing
now i check if it actually does the job
1
1
54