Rewear

Plan outfits from clothes you already own, and think twice before buying more.

Live: https://rewear-henna.vercel.app (no login)

Rewear is a small AI-assisted outfit planner. It is a prototype for UNSW INFS5706 AI in Action, where the brief is to design and test a clear boundary between what the AI does and what the human keeps.

The AI does the combinations. The human keeps taste, dress-code judgement and spending decisions.

---

The problem

Most of us wear a small share of what we own and still feel we have "nothing to wear". That feeling drives repeat purchases of things we already have. Rewear works on both ends:

1. Plan an outfit: suggestions built only from your own wardrobe, for an occasion, a dress code and an optional venue. The temperature is filled from a forecast for your area (or the venue) at the time you choose, and stays editable.
2. Talk it through with Hem: a short conversation with Rewear's AI stylist to refine the looks ("something warmer", "different shoes").
3. Wardrobe: search, filter, edit and remove items.
4. Add items: photograph clothes (several at once) and the AI drafts their tags, which you confirm one by one.
5. Before you buy: see which things you already own play the same role as something you're considering, and roughly how many new outfits it would really add.
6. What's in style: trends summarised from recent Australian fashion media, each linked to its source articles, with ways to wear them using what you already own. A reading list, not a shopping list.

The human–AI boundary

AI may	Draft tags (type, colour, formality, warmth, notes) from a photo of one garment. Combine catalogue items into up to 3 outfits with a one-line reason, using the venue type and forecast as context. Refine looks in conversation as Hem, a clearly labelled AI persona. Show overlap between a possible purchase and the wardrobe, with an estimate of new combinations. Summarise trends only from the supplied articles and map them to owned items.
AI must not	Use items outside the catalogue. Save anything. Recommend buying or not buying, or comment on price. Give body-shape, size, fit or "flattering" advice. Suggest shopping when the wardrobe falls short. Claim to be human. Invent trends, brands or sources.
AI hands to the human when	The dress code is "not sure" or ambiguous, or the occasion is high-stakes (interview, formal event): a Your call note says what to check. A photo isn't a single garment: no tags, just the reason. Every purchase check ends with "This is information, not a recommendation. Your call." (fixed UI text, not AI-generated). The venue type suggests a different formality from the chosen dress code: the dress code wins and a Your call note asks them to check.
Human then	Checks and edits every drafted tag field, and nothing saves without Save. Edited fields are marked · edited. Picks the venue from search results, and can overwrite the forecast temperature (their number wins). Chooses which outfit, if any, to wear. Decides whether to buy, and whether a trend is for them.
How the user knows AI is involved	Everything AI-generated carries an oxblood label (AI suggestion, AI tags — check before saving). Oxblood is used for nothing else. The prompt version is in the footer, a session log lists every AI call, and an About dialog explains it all.

How it works

Browser (components/, app/*/page.tsx)
  │  catalogue (seed JSON + this browser's additions) and form inputs
  │  photos re-encoded on a canvas first: EXIF/GPS stripped, ≤1024px, never stored on a server
  ▼
/api/plan · /api/stylist · /api/tag · /api/buy · /api/trends   (AI routes)
  └─ lib/runFeature.ts: validate → daily cap → versioned prompt → Gemini → parse
/api/locate · /api/venue · /api/forecast   (no AI: lib/place.ts)
/api/trends also reads lib/feeds.ts (fashion media RSS) before calling the AI
       ├─ lib/prompts.ts   prompts/<feature>_<version>.md, bundled at build time
       ├─ lib/gemini.ts    the only module that calls the API (key is server-side only)
       └─ lib/parse.ts     all AI-response parsing in one place

• Stack: Next.js (App Router) + TypeScript, Tailwind v4 (design tokens in app/globals.css), @google/genai. No database, no accounts.
• Model: gemini-3.5-flash-lite, set in one constant in lib/gemini.ts. During the build, gemini-3.8-flash returned constant 503 "high demand" errors on the free tier, and gemini-3.6-flash allowed only 20 requests a day; Flash-Lite allows about 500. Transient 500/503 errors are retried twice. When Gemini's per-minute limit is hit, the photo queue waits and retries; when the daily quota is used up, the app says when it resets.
• UI and logic are separate. Components only render. Everything else is in lib/ and the API routes, so the front end can be redesigned without touching AI logic.
• Storage: the seed catalogue ships as data/catalogue.json. Items you add live in your browser's localStorage (rewear.added.v1); edits and removals of shipped items are kept separately (rewear.edits.v1). Export catalogue (JSON) downloads the full wardrobe in the seed format, and Reset to my catalogue undoes all local changes.
• Outfit pictures: outfits show each garment, using a small photo the person chose to keep on their device only (IndexedDB, ~480px, re-encoded, never uploaded) or a colour silhouette drawn from the item's type, notes and colour tag. This replaces the original "never store photos" rule with an opt-out, device-only copy. The catalogue export doesn't include photos.
• Context data (no AI, no keys): approximate area from Vercel's IP geolocation headers (not stored); venues from OpenStreetMap Nominatim (© OpenStreetMap contributors, ≤1 request/second, searched on demand only); hourly forecasts from Open-Meteo (CC BY 4.0). All are called from the server, so they never see the viewer's IP.
• Trend sources: public RSS feeds of ELLE Australia (fashion), marie claire Australia (fashion) and Who What Wear Australia, cached for 3 hours. Only articles from the last 30 days are used, the page links to every source, and article links are attached by the server (the AI only cites article IDs). Trend results are cached in the viewer's browser for 12 hours to save quota.
• Abuse limit: a best-effort in-memory daily cap per IP and globally (RATE_LIMIT_PER_IP, default 80; , default 200). It's per server instance; cost safety comes from using a free-tier key with no billing attached.

Evidence features

Versioned prompts. Each feature's system prompt is its own file in prompts/: plan_v1.md, plan_v2.md (adds venue, time and forecast context), stylist_v1.md (Hem's persona and rules), tag_v1.md, buy_v1.md and trends_v1.md. The active version is set in prompts/config.json:

{ "plan": "v2", "tag": "v1", "buy": "v1", "stylist": "v1", "trends": "v1" }

To test a redesign, add a new version file (e.g. prompts/plan_v3.md) and change one line in the config. The build fails if the config points to a file that doesn't exist. The active version shows in the footer (Prompt: plan v1) and is recorded with every call.

Session log. The footer's Session log panel lists every AI call in the current browser tab: timestamp, feature, prompt version, model, status, the inputs (never images) and the AI's raw output, including failures and responses that couldn't be parsed. Download log as CSV exports it for the test log. The log stays in sessionStorage and is never sent anywhere.

Run locally

Requires Node 20+ and a Gemini API key from Google AI Studio.

npm install
cp .env.example .env.local   # then set GEMINI_API_KEY=...
npm run dev                  # http://localhost:3000

npm run build bundles the prompts and makes a production build.

Deploy (Vercel)

1. Import the GitHub repo in Vercel. The framework is detected automatically.
2. Add the environment variable GEMINI_API_KEY (Production and Preview). Optionally add RATE_LIMIT_PER_IP and RATE_LIMIT_GLOBAL.
3. Deploy. Every push to main redeploys.

Test results

To be completed. Test cases, inputs, expected and actual behaviour, with session-log CSV excerpts.

#	Feature	Input	Expected	Actual	Pass?

Redesign (before/after)

To be completed. What failed in testing, what changed (prompt version, code), and the result.

Issue found	Before (v1)	Change	After	Evidence

---

Student prototype for UNSW INFS5706 AI in Action. Uses a third-party AI API (Google Gemini); on the free tier, Google may use submitted content to improve its services. Suggestions can be wrong. AI-assisted build: made with Claude Code.