跳到正文
DEV Community· Yoonji Kim·· 6 小時前AI 評分62

AI 助理代辦生活雜務第 16 天:所需數字散落在不同地方

Day 16: Every Number I Chased Was in a Different Place

AI 導讀

Yoonji Kim 記錄把生活雜務交給 AI 助理第 16 天的經驗,發現購物、交通和監察所需數字分散在不同地點、渠道、日期及表格欄位。文章整理三組可複用提示詞,分別用於特賣場比價、完整行程搜尋及定時監察。作者亦記錄一次把朋友的行程誤認為自己的假設,修正後重新搜尋航線,並指出全程沒有購物、訂票或代她發訊息。

正文

"It's -5 °C, this is insane. I need to buy a coat or a padding fast. But — is a puffy padding actually warmer than a coat style?"

07:40 on a Saturday. I hadn't opened my laptop. I was standing in front of a wardrobe with nothing warm enough in it, and the temperature outside had just done something rude.

What followed was eleven hours of numbers. A down ratio. A hall price. A fare. A sale that didn't open. A points total. And by evening I'd noticed something that took me sixteen days to see clearly:

Almost none of the numbers I needed were on the price tag. They were in a different building, a different channel, a different day of the week, a different field of a form.

This is the log of that Saturday — day 16 of handing my life admin to an AI assistant — with the three prompts at the bottom that did most of the work.


1. Warmth is not a silhouette

I asked the obvious question: puffer or coat?

The answer isn't "puffer". Warmth comes from fill weight and from how well the neck and hem stop draughts. But a puffer tends to win in practice for a boring reason: coat-style paddings often cut fill to keep the line slim, and a wide V-neck lets cold air straight down your chest. So the honest answer was "both can work, and the one that keeps its fill and closes at the neck is warmer than the one that looks thicker".

That's not shoppable. So the second question was the one that mattered: does it exist in my size?

My size is at the large end of what Korean womenswear stocks (95–100). The list I'd been carrying around for a week had six names on it. The audit came back:

Candidate Price seen Verdict
National Geographic Cayman Pro Long N254UDW660 ₩233,550 (coupon) / ₩307,300 goose 80:20, matte, up to 100 — buyable
Discovery Leicester G DMDJ67056-BKS ₩307,300 goose, matte, 95–110 — buyable
Black Yak 3BYPAWU901BK ₩91,740 duck 80:20, stock unverified
Nice Claup A254PWDB17 ₩309,000 only 55–66
Carries Note CNEWDWA04BK ₩201,660 only 66 (≈95)
List TSDJPP82010 ₩341,140+ 66/77 stock unverified

Six names, two actually purchasable in my size, both matte black goose down and both carrying photos I could look at. Two of the six were eliminated by a single field I would never have checked on my own: the size dropdown.

I wrote about this on day 12 — N candidates existed, M were checkable, K are buyable under my conditions — and this time I didn't have to ask for it. It came back as a table with the fails included. That's the part that feels different after two weeks: the rule survived without me repeating it.


2. The outlet question wasn't about price. It was about which hall.

Then I said something that turned out to be a real question:

"I'm thinking about Paju Premium Outlet. What's the price range there? When I look on Naver Fashion Town, it doesn't look much different from the other malls."

I was right, and that was the interesting part. The outlet wasn't lying, and Naver wasn't lying. The discount lives in a different building inside the same complex.

  • Main stores carry the usual 30–50% outlet discount. Against an online coupon, that's a rounding error. Hence "no difference".
  • The real 50–80% sits in pop-up and special-event halls, for a few days only.

For that week, the Paju special hall (11/19–25) listed:

  • Coats, She's Miss / List: ₩109,000–119,000
  • Duck-down paddings: ₩159,000–169,000 — less than half of the same brand's ₩341,140 online price
  • Eider up to 60%

And then the thing that decides whether you actually go: the clock.

Place From my neighbourhood Closes
NC Bulgwang 7 min —
Hyundai City Outlet Dongdaemun 26 min 21:30
Lotte Seoul Station 28 min —
Lotte Premium Outlets Paju 58 min (GTX-A) 21:00
SPACE1 Namyangju 66 min —
Mario / Hyundai Gasan 73 min 21:30
Hyundai Gimpo 94 min 20:00

At 17:25, Paju was still reachable and Dongdaemun was still comfortable. Everything past 70 minutes was a gamble on traffic.

The honest trade-off I put in front of myself: ₩110,000 coat, if it exists in my size, in a hall that closes at nine versus ₩233,550 coat, definitely in my size, delivered. A special hall is a lottery with good odds. Online is not a lottery at all. I didn't go that evening — the weather was for a long padding, but I don't own one, and the coat I'd wear to something nice is still waiting to be dry-cleaned. Both are missing, and "both are missing" is not a shopping answer, so I left it for tomorrow.

What I learned: "30% off" describes nothing until you say which building. Ask for the hall, the dates, and the closing hour — the three things a price comparison site never shows you.


3. A fare is a channel before it is a price

Then a one-line message arrived that changed a rule permanently:

"$117 is how much. All foreign currency prices should be shown with the won too. I already did the bus. Track Fukuoka→Sapporo on 1/8 and 1/9 too — train, bus, flight. Doesn't have to be direct. Connection combinations too."

$117 at 1,346 won to the dollar is ₩157,482. The same one-way was ₩157,638 on one aggregator and ₩153,940 with a card discount on another. Three numbers for one seat, and I'd been quoting the one in dollars as if it settled something.

So the sweep ran. Fukuoka to Sapporo, 1/8, per person, one way:

Option Time Total
Jetstar direct 14:35→16:55 2h20 ₩118,900 (₩98,814 before fees, airline site)
Peach direct 11:10→13:30 2h20 ₩118,600
One stop via Tokyo 4–7h ₩157,000+
Two separate LCC tickets via Narita +4h ₩167,363
Shinkansen + Hayabusa + Hokuto 15h40 ₩500,300
Night bus + rail ~26h ₩371,300
Ferry routes (2 nights at sea) ~28h ₩195,650 — not available on those dates

Three things in that table are worth more than the table:

The channel changes the price more than the airline does. On the Sapporo→Okinawa leg the same Jetstar ticket was ₩129,224 on Jetstar's own site, ₩138,890 on a Japanese aggregator, and ₩192,900 on the Korean one. The airline's own checkout was cheapest, twice, by ₩8,000–9,700. Booking through an aggregator is a convenience fee you pay without being told.

A holiday is a 100% surcharge. Peach was ₩118,600 on 1/8 and ₩236,200 on 1/9 — because 1/9–11 is the Coming of Age Day long weekend. A fare table without the calendar in it is a trap.

Ground transport was never competitive. Fifteen hours and ₩500,000 to do by rail what ₩118,900 does in two hours. I still checked, because "doesn't have to be direct" was an instruction, not a preference — and now the answer is closed rather than assumed.


4. The correction: the itinerary is mine, not the tracker's

Here's where I got it wrong, and it's the most useful failure of the day.

I had built a tracker. It had the right segments. But when I answered, I asked a question that revealed my assumption — whether to switch tracking to Sapporo→Okinawa on 1/10, since that would put me in Okinawa on the 10th "rather than the 8th".

The reply was four words and a comma:

"No — my friend's plan is Fukuoka→Sapporo by train, Sapporo→Okinawa arriving the 10th."

The segments were right. The person was wrong. I had silently attached the itinerary to the wrong traveller, and then built a recommendation on top of it.

After the correction, the 1/10 Sapporo→Okinawa search had a different shape:

  • Direct: exactly one flight. Peach MM273 11:30→15:40, ¥19,030 = ₩163,658, 7 seats left.
  • ANA's old weekly direct (NH1693) is absent from the winter schedule — and Okinawa's small airport has no Sapporo service at all. So "check ANA" is not a search, it's a dead end, and now it's written down as one.
  • Cheapest overall: Jetstar single ticket via Narita, 12:00→19:15, ¥15,026 = ₩129,224, 7h15 total with a 1h50 connection.
  • Thirteen other combinations were priced and all were more expensive. Separate tickets never beat the single one, and none of them protect you if the first leg is late.

What I learned: when a request has a second person in it, the first deliverable is the sentence "here is the route I think you mean" — not the price table. A wrong assumption doesn't show up as an error. It shows up as a confident answer to a question nobody asked.


5. A sale that doesn't open is also information

There was a Black Friday sale expected at 12:00. So at 11:46 I started checking: the Japanese site, the Korean site, the notice board, the press page, twelve banner links. Nothing at 12:00. Nothing at 12:18.

The useful part wasn't "no sale". It was this: all nineteen of that airline's past sales had started on a weekday at noon. So the absence wasn't a failure of the check — it was a schedule fact. Next candidates: 11/24–27, and the tracker now waits for those instead of refreshing a page that has already answered.

And then the thing I didn't expect to find while looking at a sale page: a data-breach notice, dated the day before, covering 8,997,106 customer records leaked through the airline's email vendor — names, addresses, itineraries. I had a booking on that airline from two weeks ago.

Would I have seen it on my own? Maybe, in a news cycle, in a week. I was in the account, in the notice board, for a different reason. Standing in someone's account for one purpose makes you responsible for what else is on the wall. It went into the same message as the sale update, one line, no drama — which is how I want to be told.


6. Twenty-two checks that found nothing

All day, every thirty minutes, a routine opened a ticket site and looked at one block of seats: front rows, middle seats, in a show I'm seeing in December. I already hold a seat two rows further back, bought with free cancellation to 11/27. If a better one frees up, I'd rather know before that window closes.

Twenty-two runs between 05:43 and 16:43. Active seats on sale drifted down over the day — 525, 520, 519, 512, 511, 506, 499, 496, 494, 491, 487, 484, 481, 479 — and zero of them were in my block.

That's the least glamorous shape of delegated work: nothing happened, twenty-two times, and I only ever saw a message when there was something to see. It costs a check; it buys a season of not wondering.


7. And then my own ledger broke

At 12:30 the points check came back at 1,473 — verified on the site, not from memory. But the same run found that the file I use to keep that count had a missing comma and a duplicated key in it. It still parsed. It would not have parsed next time.

I'd been treating that file as "my" memory. It's a text file written by a script under a deadline, which means it can be wrong in exactly the way a spreadsheet can be wrong: quietly, and only later.

Fixed, re-verified, and then the midnight run added 12 more — attendance plus two comments — taking it to 1,485. Fifteen points from the next tier, with a projection that doesn't need luck: attendance and comments alone cross the line on 11-24, or tomorrow if one handwritten transcription photo arrives.

What I learned: verify the thing that keeps the score, not just the score. A ledger nobody audits is a rumour with formatting.


8. The 17:14 that didn't run

At 17:14 the seat check failed — not on the site, but before it started: the weekly usage allowance behind it was used up. No new task could launch until it refilled.

I was told in the same minute. Not "I'll check later" — the actual reason, so I could decide whether it mattered. My answer:

"Skip the ticket check today. What about coat special-event halls or other offline shops? Tomorrow is padding weather, but I haven't sent my nice coat to the dry cleaner and I don't own a long padding."

So the monitor was paused, and — this is the part I'd have forgotten two weeks ago — an automatic resume was registered for 08:00 the next morning. A pause without an end time is a silent cancellation, and a silent cancellation is how a routine dies: nobody decides to stop it, it just stops being true.


9. A review you can stand behind

Late afternoon, a photo from August came up in the blog backlog: a small restaurant near my place, dinner at 9pm.

"This is Seohyeonine Sikdang. I won't go back, but I can't write badly about it either — how should I put it? I should crop my hand out of the photo. The flavour was flat. Because my guest was foreign, they quietly cooked it without chilli and offered chilli on the side, but the seasoning was bland."

This is the one I found hardest all day, and it isn't a technical problem.

The facts, in order: a plain home-style restaurant; a foreign guest; the kitchen took the chilli out without being asked and offered it separately; the result was under-seasoned. None of those is a complaint. All of them are true.

So the framing became: "not spicy, plain home-style cooking — and considerate about chilli when you bring a guest who doesn't eat it." No rating, no verdict, no "disappointing". A reader who wants heat reads "plain" and decides for themselves. The hand gets cropped. That's it.

What I learned: an honest review is the facts you'd tell a friend, minus the adjective that settles the argument. "I won't go back" is mine to keep; "the seasoning was bland" is theirs to read.


One Saturday, in numbers

22 seat checks, 0 seats found in the block
6 → 2 outerwear candidates → actually buyable in my size
9 transport options compared door to door
3 prices for one fare, depending on the channel
1 itinerary assumption corrected
1 ledger bug found before it mattered
15 points to the next tier
-6 °C tomorrow's low, and the reason for all of it

Three prompts you can copy

1. The "which building" prompt — for any outlet or sale question

I'm considering [outlet / mall / sale] for [item]. Before any prices:
1) Tell me which parts of it discount differently — main stores,
   pop-up halls, special-event halls — and what % each actually gives.
2) Give me the dates each event runs, and the closing time that day.
3) Quote prices only from the event halls, with the brand and the
   original price next to it.
4) Then the trade-off, plainly: what I can definitely buy in my size
   online at [price], versus what I'd be gambling on finding there.
Ask me nothing until you've done 1–3.

Why it works: it forces the answer to name a location and a date. "Up to 70% off" is marketing; "She's Miss coats ₩109,000–119,000 in the pop-up hall, 11/19–25, closes 21:00" is a decision.

2. The full-route sweep — for any leg of a trip

Find every way to get from [A] to [B] on [date], for one person, one way.
Include: direct flights, one-stop flights on a single ticket, separately
ticketed connections, rail, overnight bus, ferry — even the ones that
look silly.
For each: departure→arrival, total duration including connections, total
price including all fees, and what baggage is in that price.
Rules:
- Show the local currency and my currency for every price.
- Note which fares are impossible on that date rather than leaving them out.
- Tell me the holiday calendar around that date and whether fares jump.
- For the cheapest three, say which channel you priced it on (airline
  site vs. aggregator) and check the same ticket on the airline's own site.
End with one line: cheapest overall, cheapest direct, and what the
difference buys me.

Why it works: the channel line alone saved ₩9,700 on one ticket. The holiday line explained a ₩117,600 jump that would otherwise have looked like a pricing error.

3. The monitor prompt — for anything you'd otherwise refresh

Watch [thing] for me: [exact block / exact item], checking every
[n] minutes while [condition].
Do without asking: look, compare, record what you saw.
Never: buy, hold, or log in on my behalf beyond looking.
Report only when [condition is met]. Otherwise stay silent — I don't want
a message per check.
Each check, note the total count of available items so we can see the
trend even when mine never appears.
If a check can't run (technically or because you've hit a limit), tell me
then, with the reason, in one line.
If I say skip / pause: pause it, and register the time it resumes by
itself. Never let a pause become the end.
If, while looking, you find something else that affects me — a notice,
a breach, a deadline — tell me in the same message, briefly.

Why it works: it defines silence as a valid result, gives "nothing" a number to move, and makes a pause reversible without me remembering to undo it.


What didn't change

The week's rules held without being restated: nothing bought, nothing booked, no account created, no message sent in my name. The seat check still stops at the seat map. The ticket stays unpurchased while the price is still moving. The coat is still unchosen, which is a truthful state, not a failure — I have a wardrobe problem, not an information problem, and the information part is now done.

Sixteen days in, the thing I actually rely on isn't the searching. It's the two sentences I get for free: "here's what I assumed" and "here's what I couldn't check." The numbers were everywhere this Saturday — in a hall, in a checkout, in a calendar, in a size dropdown. None of them were where the price tag said they'd be.


Korean version of this post: 티스토리

來源:DEV Community · dev.to