Built for the Kiroween 2026 hackathon (Resurrection category). #KiroweenHackathon26
In the mid-1980s a company called Cauzin sold the Softstrip: a reader you swept across strips of dots printed in magazines to load software into your computer. Paper was the network. It disappeared, but the idea didn't.
Strip is that idea rebuilt for the phone already in your pocket. Pick a file and Strip turns it into printed pages of codes, or an animated stream on screen. A phone camera reads it back and rebuilds the file. It runs entirely in the browser, works offline, and nothing is uploaded.
- Live: https://strip-seven.vercel.app
- Code (MIT): https://github.com/alwayskumar-global/strip
How it works
- Fountain code. The file is split into blocks and encoded with a systematic LT fountain code (robust soliton distribution). Packets can arrive in any order and some can be lost, and the receiver still recovers the file. That matters for paper (a crease, a smudge) and for a camera watching a screen (dropped frames).
- Packets to symbols. Each packet is CRC-checked and carried in a QR symbol. Pages hold 9 symbols each, with repair pages built in so no single page is critical.
- Decode in a worker. The receiver uses a self-hosted WebAssembly decoder (zxing-wasm) in a Web Worker, feeds packets to the codec, and verifies the final file with SHA-256 before offering the download.
-
Strict by default. CSP with no inline styles,
connect-src 'self', camera permission limited to the app. Tests assert that no network request leaves the device and that storage stays empty.
What Kiro was good at
The whole repo was written in Kiro, spec-first, as three specs: codec, symbols and app. Each one went requirements → design → tasks, with numbered correctness properties, property-based tests (fast-check) and one commit per sub-task.
The spec process caught my own wrong claim. I had written a recovery target: decode from about 1.1× the packets needed. During design, Kiro computed that this is mathematically unreachable for a plain LT code at these sizes. Instead of quietly weakening the test, we changed the requirement: three measured claims (late join, cold start with 30% loss, paper mode at 5% loss), each with trial counts and Wilson confidence bounds, plus dense repair rows found through a parameter sweep. The published overhead table is measured, not hoped for.
It made me pick a decoder with data. The first QR decoder failed my own decision rule on rotated and skewed photos (0.87 success vs 0.95+ for zxing-wasm). Kiro ran the spike on the same degraded images, wrote it up in docs/spikes/, and we switched.
Property tests found real bugs. An encoder that threw a TypeError on null input, a spec conflict about which rejections count at packet level, and a log-bounding counterexample all came from properties, not from example tests.
Stop-and-wait with credit caps. I gave each task group a credit cap and made Kiro stop and report at every boundary. It kept a 4,000-credit budget predictable, and the reports made review quick.
Design fidelity as a test. The 34 UI boards were designed outside Kiro in Claude Design, and Kiro converted them into components. A fidelity harness renders the originals and the app and diffs pixels plus DOM structure (53 targets). It caught a primary button that was invisible because of anchor specificity, a bug I'd never have spotted by eye.
Numbers
- Codec: 316 tests. Symbols: 148. App unit: 132. Tooling: 94. Playwright: 233, including full round trips (file → printed pages → page images → byte-identical file).
- First load: 179.6 KB gzip against a 400 KB budget; the WebAssembly decoder loads separately.
- Decode time: 3.6 ms per 720p frame, 8 ms at 1080p.
- A Samsung Galaxy S23 received a 90 KB image from a laptop screen in a few seconds. [UPDATE with the printed-page result, or delete this line if the paper test isn't done.]
Honest limits
- Files up to 256 KB; paper gets unwieldy above 64 KB.
- Above roughly 1,000 blocks, the Gaussian-elimination cap makes late-join recovery slower. The README's table shows the cost per size.
- Animated QR file transfer already exists. Decimen is the closest relative, and TXQR, libcimbar, PaperBack and qr-backup came before. Strip's angle is the print path and the Softstrip story, with a measured codec behind both.
What I'd do next
Compression before encoding, optional encryption, and a lot more real-device testing across phones and printers.
If you want to try it, open the live link, drop in a file, and point a second device at it. Feedback and issues are welcome on the repo.
Top comments (0)