Backpocket
Backpocket is a WhatsApp contact you share links, screenshots and reels to. It turns every save into a structured card (title, summary, key points, topics), stores it, and sends the right one back when you ask for it in the same chat.
Why I built it
I save things by sending them to my own WhatsApp chat: LinkedIn posts before coffee chats, recipes, how-to reels, internship postings. I almost never open them again, and when I need one I can't find it. The problem isn't reading. It's timing: a saved post is useful on the morning of the chat it relates to, not the day I saved it.
So Backpocket lives where the saving habit already is. You share to it like any contact, ask it questions in plain language, and tell it what you acted on ("applied to 1 and 3").
What works today
- Save: links (LinkedIn, recipe sites, articles), screenshots and plain notes become searchable cards.
- Find: "what did Aditi say about growth cases?" gets a short answer that quotes the original message you saved.
- Lists: "share all summer 2027 internship links" returns a numbered list; reply with numbers to mark them applied, cooked or read, and they drop out of future lists.
- Morning brief: at 6am, one message with what to read before today's meetings, the job posts you saved but haven't applied to, and any applications it found in your email. Quiet days get nothing. Setup is in bridge/README.md.
Not built yet: Notability notes, and reading application emails in a UNC Outlook inbox.
Product decisions
| Decision | Alternative I rejected | Why |
|---|
| WhatsApp as the only interface | Telegram bot, web app | It's where I already save things. Telegram was an hour less setup but a habit I don't have. |
| Ask for a screenshot when LinkedIn or Instagram hides a post | Scraping behind logins, downloading reels | Scraping breaks the platforms' terms and fails silently. A screenshot works every time and the user knows why it's asked. |
| Claude decides what a link-free message means, with a keyword check as backup | Keyword rules only | "Done with Nordstrom" or just "2" only makes sense next to the last list sent, which keyword rules can't see. The backup keeps the bot answering when the model call fails. |
| Answers capped near 200 characters | Full summaries | Early answers restated everything in the saved item. A reply should read like a friend who read it for you. |
| Strip tracking parameters before saving links | Store links as shared | The same LinkedIn post shared twice had different URLs, so duplicates slipped through. |
| Calendar and inbox read by an Apps Script in my own Google account | "Sign in with Google" | Google requires a paid security assessment for inbox access, and without it the login expires every 7 days. The script keeps the inbox inside Google and sends only confirmation-email metadata. It doesn't scale past a handful of users, which is fine for now. |
| No brief on quiet days | A daily message no matter what | A brief that's empty half the time teaches you to ignore it. |
| Remind about a saved job at most every 3 days, sooner near its deadline | Every unapplied post, every morning | Nagging is how reminders get muted. |
How I built it
I wrote the spec and the rules in CLAUDE.md and built with Claude Code. Every change after the first version came from using it on my own saves: two screenshots racing for one link, LinkedIn previews cutting posts short, answers running too long. The commit history shows that trail.
Architecture
- WhatsApp Cloud API is the only interface. Meta calls
POST /api/whatsapp for every incoming message.
- Next.js route handler on Vercel verifies the
X-Hub-Signature-256 signature, returns 200 right away, and does the work in after() so Meta never retries a slow webhook.
- Extraction (
lib/extract.ts) reads Open Graph tags, page titles and JSON-LD recipes from links, and downloads screenshots. Pages that hide content behind a login are marked needs_fallback and the user is asked for a screenshot. No login scraping, no video downloads.
- Claude (
claude-sonnet-5-5) turns the extracted content into a card using a strict tool schema, and picks the best matches when you search.
- OpenAI
text-embedding-3-small embeds each card for semantic search.
- Supabase stores cards in Postgres with pgvector (
items, resurfacings, match_items()), and screenshots in a private Storage bucket.
- Morning brief (
lib/brief.ts): a Google Apps Script (bridge/) posts today's events and confirmation emails to POST /api/brief, signed with BRIDGE_SECRET. The server adds the published Outlook calendar if there is one, matches emails to saved job posts with Claude, embeds each meeting to find related saves, asks Claude which ones genuinely help, and sends one message. If WhatsApp's 24-hour window has closed, the brief waits in chat_state until the user writes.
Setup
- Install Node 20 or newer, then run
npm install.
- Create a Supabase project. In the SQL editor, run the files in
supabase/migrations/ in order.
- In Supabase Storage, create a private bucket named
screenshots.
- Copy
.env.example to .env.local and fill in every value. The app names any missing variable when it starts.
- Run
npm run dev and open http://localhost:3000/api/health. You should see {"ok":true}.
Scripts
| Command | What it does |
|---|
npm run dev | Local dev server on port 3000 |
npm run build | Production build |
npm run lint | ESLint |
npm test | Vitest unit tests |
npm run simulate | Sends a signed fake WhatsApp webhook |
npm run simulate-brief | Builds the morning brief for a sample day (add --send to send it) |
npm run backfill-jobs | Finds job posts among items saved before job detection |
Simulator
scripts/simulate.ts builds a realistic WhatsApp webhook payload, signs it with WHATSAPP_APP_SECRET from .env.local, and POSTs it to the webhook. It lets you test the full pipeline without a phone.
npm run simulate -- --text "https://www.example.com/some-article"
npm run simulate -- --text "what can I make with paneer?"
npm run simulate -- --text "hello" https://your-app.vercel.app/api/whatsapp
The URL defaults to http://localhost:3000/api/whatsapp. The sender is the first number in ALLOWED_PHONES.
Against a server with a real WHATSAPP_TOKEN (including the deployed app), every simulated message gets a real WhatsApp reply on that phone, and test items mix with real ones. To test without sending anything, run the local server with a fake token: WHATSAPP_TOKEN=fake npm run dev.
