Ctrl F for Screenshots

A small tool I built for myself went from 2 stars to 300 in four days, then I spent the rest of the week on a round screen the size of a watch face.

weekly 017Oct 11, 20264 min read

I built a thing to solve a problem I actually had. It turns out a lot of people have the same one.

gyotaku

I screenshot everything and can never find anything. So I made Ctrl F for my screenshots folder: type any word that was in the image, it shows up in milliseconds, fully offline.

Four days earlier it had 2 stars. One launch post later it was at 289 stars and 1,724 downloads. It's at 309 now. The post did 48,000 views.

I also put it on Show HN, which is where most of the serious questions came from.

Then I spent three days shipping everything people asked for. Six releases in about 72 hours.

It runs on Windows now, packaged by CI with onnx runtime beside it, one line in PowerShell. Near-match search, so a word the OCR half-misread still finds the screenshot. Clipboard images get saved and read. Devanagari, so it reads Hindi, Marathi and Nepali, plus vertical Japanese. Every shortcut rebindable. A mark-and-trash flow with one undo, because Harshit said he takes screenshots for one-time use and never deletes them.

And it runs on macOS, which I did not do: saurav-codes ported the search window and background reader to Apple Silicon, Chiroyce1 made it default to Cmd shortcuts there, and RawJat fixed the search window opening on the wrong monitor. Three strangers, one of them shipping a whole platform.

It has a home at gyotaku.app now, with a landing page built in the gyotaku print style on Hokusai and Hiroshige backgrounds.

The name is a Japanese printing technique, and the mark is a fish printed in lines of text.

Breakscale, in Claude

Breakscale is live in Claude's connector directory. You can ask Claude to design a system and then actually run it.

I asked for a URL shortener at 2,000 requests a second: 89.4% of requests failed with postgres pinned at 100%. Then "fix it with the fewest changes", and it made one change and re-ran it.

The same day, Workers IO launched publicly, which is the thing I'm building with Chaitanya. The pitch is millions of simulated futures for your actual system, and every failure replays exactly. Breakscale is at 1,376 stars.

ESP32 week

A friend handed me a round 1.75 inch AMOLED board with touch, a speaker, a mic and a battery. I lost the weekend to it.

First roundkit, the shared layer everything else sits on: battery and charging from the AXP2101, tilt and gyro from the QMI8658, tap and long-press gestures, saved settings, a pixel surface with dirty-box tracking so far-apart updates stay small, and a Docker build so none of it needs a local ESP-IDF.

Then five things on top of it, all C++:

cassette-orb: joins wifi, polls my site, and shows whatever I'm playing with a cover, a halo, scrolling text and an equalizer.
claude-orb: a pair of eyes that reacts to my Claude Code sessions.
hourglass: a pomodoro timer where you flip the board and the sand really falls.
shaza: tap the screen and it tells you what song is playing. The simulator uses the laptop mic.
xevrion-card: a round digital ID, swipeable pages, scannable QR, fully offline.

A quieter site

Rewrote most of xevrion.dev around a DESIGN.md, a taste file that every change now has to follow. Tokens, type scale, motion, the accessibility floor, the voice. Having the rules written down made every later decision faster.

New dock that becomes whatever you type, say-hi messages with daily and site-wide limits, live star counts, Paper Mono for labels and counts, a home page rendered complete on the server, llms.txt and security.txt.

Also removed the goals page. It had no real use, and I don't want everything about me public.

Which is the other thing: I'm building something paid. It's for anyone who has ever launched something and got three likes. Not saying more yet.

What I found interesting

fsearch : whole-disk file search for macOS, about 1ms across 8 million files. Uncomfortably adjacent to my week.
rea : reverse engineer anything with agents, from app behaviour down to native binaries.
emilkowalski/skills : skills for designers and engineers.
OmniWM : tiling window manager for Apple Silicon with Niri-style scrolling containers.
snapshift : Anish's chat that reads the whole conversation every turn and picks the UI for each answer.
cognee : open-source memory platform for agents.
iina, SimpMusic and freedoom : a good video player, a YouTube Music client, and a completely free Doom.

Random

2,000 followers. End of August it was 1,100. What got me the rest was shipping things and posting the video of them, which is not a complicated strategy.

Also turned up as a GSoC Preferred Candidate, which means Google shares your name with their hiring team after you complete the program. Didn't know that was a thing.