AI-native product · Personal project · Native iOS app · October 2026

Snapee

Take two photos and, a minute later, you have a price you can explain and a listing you can post. Snapee (一拍即出, “snap and it’s out”) is an idea that came after my master’s capstone: watching Goodwill’s processors price second-hand clothes, and remembering the things I couldn’t sell when I moved. I built it as a native iOS app with a real vision model, tested it on my own clothes, and gave it a design system of its own.

Three iPhones: step one, “photograph the whole garment”, with a coral jacket mascot raising both sleeves, a blank tag hanging from one cuff; the cream hang tag while analyzing, with brand, size, material and condition still reading “recognizing”; and the result, a hang tag saying this red top should sell for ¥65, with a slider from “sell fast ¥50” to “hold out ¥80”
Three screens: capture, analyzing, result. One cream hang tag runs through the flow: it appears while analyzing and ends with the price on it. The mascot is a coral jacket with a blank tag on its cuff. The product is in Chinese; captions translate what matters.
Origin
A personal project after my master’s capstoneThe capstone solved second-hand pricing with Goodwill; this project brings the same problem back to ordinary people.
Timeline
October 2026
My role
EverythingProduct definition · design system and screens (Figma) · model prompts · walkthroughs, error-handling tests and the real-clothes test · iOS development, with an AI coding tool
Stack
SwiftUI · Doubao vision model (Volcengine Ark) · code written with an AI coding tool

At a glance

0 → 1From idea to a native iOS app, by one person
1Design system of its own: ivory canvas, coral, a cream hang tag and a jacket mascot
2 callsTwo model calls per garment: one to read it, one to price it and write the listing
3 of 3Real-clothes test: brand, size and material all read correctly from the label

Problem

01 / 04

Selling what you own gets stuck on the price

Every time I move, I get stuck on the same thing: I don’t know what to charge for the things I’m selling. Price low and it sells fast, but I feel I gave it away. Price high and it sits there, and in the end I give it away anyway or throw it out. Two things are stuck: there is nothing to base a price on, and writing the listing is a chore.

The users are people with a pile of clothes they no longer wear and no appetite for researching the market: graduates moving out, people changing cities, anyone clearing a wardrobe for the season. They want it sold, at a fair price, without much effort.

The judgment also comes from my capstone at Carnegie Mellon. We worked with the second-hand retailer Goodwill, whose processors price garments from a price sheet. Ordinary people have no such sheet. In our survey of 131 Goodwill employees, only 8% trusted an AI recommendation. So this product cannot just throw out a number. It has to show its reasons and leave the decision to the person.

Design

02 / 04

Capture → Read → Confirm → Post

Clothes, shoes and bags first. Labels have words printed on them, and what the AI reads is more reliable than what it guesses from looks. With a narrower category, the follow-up questions, the condition scale and the pricing factors can all be specific. When the photo is of something else, Snapee doesn’t force a reading. It offers “retake” and “try a sample”.

  1. 01Capture

    The whole garment first, then the label (skippable), a flaw close-up optional. From the album, one slot per photo, so the AI knows which is which.

  2. 02Read

    A hang tag appears and the AI writes what it read. The price line says “after you confirm”.

  3. 03Confirm

    What needs confirming sits on top. Every field names its source. Only three questions, all tappable.

  4. 04Post

    The tag drops in with the price. Below: how buyers will see it, two listings, the sale-card fields, one-tap copy.

Photo 2 of 2, “now the label”: the whole garment is done, with a thumbnail of a red T-shirt; the viewfinder holds the photo of the care label at the collar, marked “label done”, then a “Start reading” button and an optional “add a flaw close-up”
One thing at a time: the whole garment is done, now the label.
The pick-from-album sheet with three slots: whole garment (required), label (recommended) and flaw close-up (optional), each with a Choose button
From the album: one slot per photo, not a multi-select.
Analyzing: the heading says “Reading your garment with AI…”, the cream hang tag lists brand, size, material and condition as “recognizing…”, the price line reads “after you confirm”, and below are “usually 5–10 seconds” and the jacket mascot
Analyzing: the hang tag appears for the first time and fills in as the AI reads.

Every field names its source; only ask what a photo can’t show

The confirm page separates what the AI read from what the person has to decide. Each field carries a source label: “read from label”, “inferred from photo” or “needs your confirmation”. You see at a glance where to look twice and what you can simply trust.

“Only you know” is three questions: how often it was worn, whether there is a problem the photos don’t show, and how you want to trade. All are tappable, nothing to type. Purchase price and proof of purchase fold away under “More”. Anything left unanswered is simply left out of the listing.

Confirm page: the condition card is marked “needs your confirmation” and asks whether 90% new looks right, with “yes, 90% new” selected; below, “What the AI read” lists the item name (inferred from photo), brand GILDAN, size M and material heavy cotton (all read from label), and flaws: none visible (needs your confirmation)
What the AI read: brand, size and material each say “read from label”.
Only you know: “worn a few times” selected, “no hidden problems” selected, “local pickup” selected, and the expanded More section with a purchase price of 130
Only you know: three taps, purchase price optional.

A range, not a number, and you hold the slider

The result page puts the suggested price on the hang tag, with a slider underneath from “sell fast” to “hold out for more”. That is exactly the dilemma I have when I move, and now it belongs to the user: dragging the slider changes the price, the bargaining line in the listing and the sale card together.

“Why this price” expands into three to five reasons. Under the tag it says plainly that the estimate is the model’s judgment, not a live market price, and offers a search phrase for checking the same item on a marketplace. The listing states only facts the user confirmed, one version for Xiaohongshu and one for Xianyu.

Result page: the cream hang tag says this red top should sell for ¥65, the slider runs from sell fast ¥50 to hold out ¥80, with four chips: 90% new, worn a few times, local pickup, estimate confidence high; under the tag, the jacket mascot and “scroll down for your listing”
The tag reveals the price; the slider lives on the tag.
“Why this price” expanded with four reasons: a budget T-shirt brand, so the price sits below the purchase price after depreciation; 90% new with no flaws supports the price; bought for ¥130, second-hand at roughly 40%; worn only a few times, so the hold-out price can be a little higher
Why this price: the reasons are written on the tag.
The listing card with Xiaohongshu and Xianyu tabs: a title about a nearly-new GILDAN print tee, a body naming the brand, size, material, worn a few times, local pickup and “¥65, open to small offers”, four hashtags, and Copy title / Copy body buttons; below, the first sale-card field, the name
Two listings and the sale-card fields, one tap to copy.

All screens are the real app, walked through with one of my own T-shirts. The reading is the real model’s answer about this shirt from October 21, replayed into the app for the screenshots rather than called again.

What the AI does, and doesn’t

03 / 04

Two model calls per garment

The first call reads: one to three photos in, and out comes the category, brand, size, material and colour, each with a source and a confidence, plus visible flaws, a suggested condition grade and photo problems (blurry, not clothing, several items, no label). The second call prices and writes: the confirmed fields and the three answers in, and out come three price points with reasons, a search phrase, two listings and the sale-card name.

Everything else is local rules: the question templates, the condition scale, the slider, and validation with fallbacks (sorting the prices, filling missing fields, deleting “authentic” and “guaranteed genuine”). The slider, copying and switching platforms never call the model, so they are instant and free.

  1. Limit 01No authentication

    For commonly counterfeited brands the listing says “cannot verify authenticity, please consider before buying”, and it never says “authentic”.

  2. Limit 02No invented facts

    Purchase date, price and wear come from the user; unanswered means unwritten. A walkthrough caught a sample listing saying “worn only a few times” when the user had never been asked, so the copy is now built from the answers.

  3. Limit 03No pretending to know the market

    Only a range, with a stated confidence. Two runs on the same T-shirt gave ¥45 / 59 / 75 and ¥50 / 65 / 80. That variance is the reason for a range rather than a single number.

  4. Limit 04No retouching the photos

    Retouching can erase the flaws too, and that misleads the buyer.

When it fails, each case gets a plain sentence

Blurry photo, not clothing, timeout, no network, quota used up: each has one sentence in plain words plus “retry” and “continue with a sample”. When the photo is of something else, Snapee says what it saw instead of forcing a reading.

The app has no server yet: the phone calls the Doubao vision model directly, with a dedicated, spend-capped key created for this one project; the AI calls sit in one module, so a production version swaps in a server-side proxy by changing that module alone. Three sample garments use preset data, so someone without clothes at hand can still walk the whole flow.

“This doesn’t look like clothing”: under the jacket mascot, the AI saw something like a thermos; Snapee currently handles clothes, shoes and bags; buttons for “retake a garment photo”, “walk through with a sample” and “wrong call? It is clothing, continue”
Not clothing: say what was seen, and leave a “wrong call” path.

How it was built

04 / 04

I wrote the specs and prompts, a tool wrote the code, real clothes signed off

I wrote the product definition, checked the facts, designed in Figma (five home-screen directions → v2 → v2.1 after my own comments) and turned each round of changes into a spec for an AI coding tool to generate the code. Every round I walked the build through in the simulator. Before connecting the real model, I intercepted the model requests and simulated ten kinds of response to check every error state.

Some real turns along the way: a walkthrough found the camera button wasn’t wired to a file picker, so nobody could upload a photo; the real-clothes test found over-long reason chips and English field names leaking into the reasons. Each went into the next round.

A personality of its own

The camera and the photo library are the front door, and a half-finished garment is saved as a draft to pick up next time. The design system was built for this app alone: a warm ivory canvas, coral for the main action, a cream hang tag, a playful Chinese display face for headlines, and a mascot, a coral jacket with a blank tag hanging from its cuff. The hang tag is the motif that runs through the flow, appearing while the AI reads and ending with the price on it, and it is the cream tag from the mascot’s sleeve.

Home screen on first use: the jacket mascot raising both sleeves, the heading “photograph the whole garment first”, a coral Take photo button and a white Pick from album button, a line of photo tips and a “use a sample photo” link
First use.
Home screen with an unfinished draft: the mascot shrinks beside the heading, and under “Recently unfinished” a hang-tag-shaped card holds the red T-shirt thumbnail, “priced, ready to post” and a View listing button
With an unfinished draft.

Tested on a CMU T-shirt

Photographing the whole shirt and the care label at the collar, the model read the brand GILDAN, size M and material Heavy Cotton from the label, marked all three as “read from label”, and priced it in three tiers with reasons using the ¥130 purchase price. Reading took about four seconds, pricing and writing about four more.

A red T-shirt laid flat on a bed, printed Carnegie Mellon University
Photo 1: the whole garment.
Fingers holding open the black care label at the collar: GILDAN, HEAVY COTTON, size M
Photo 2: the care label at the collar.

What’s next

Deliberately left out: publishing straight to a platform, live price comparison, chat haggling, authentication. Three things I want to do next: run the whole flow on a real iPhone with the camera and the live model; fill the platform’s sale card directly from the reading; and widen the categories by changing the second photo from “the label” to “the one with words on it” (a nameplate, a model sticker, the box, a book’s back cover), with the same principle.

Next case study

Multi-device experience · Interaction design Social networking app In-person meetups for five, designed for Bumble BFF