# I built a laser dry-fire shooting game with Claude Code: a webcam watches your monitor and scores laser hits (open source, free) — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/e319f2dc67e9044d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-25 02:37 ingested 2026-10-26 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xeqtp1/i_built_a_laser_dryfire_shooting_game_with_claude/)

## retained text



```
I built **ScreenPlink** with **Claude Code** over about two weeks. It is a laser dry-fire shooting game where your monitor is the target. You shoot the screen with a laser trainer pistol, or with a laser cartridge in your own gun. A webcam sitting beside the monitor watches the screen rather than you, the app works out which pixel you hit, punches a bullet hole there, and the target falls.

Nine drills, fifteen figures with scoring rings, a 17-round magazine, and two players on one screen. You can also take both halves yourself with a pistol in each hand: two lasers landing in the same camera frame are detected as two separate shots.

The 90-second trailer shows what the game looks like.

The camera does not need to be in front of the screen or square to it. Off-axis at an angle is normal and supported: the app photographs its own markers on the monitor to learn the mapping, so calibration costs no ammunition.

**How Claude Code was used to build it**

Claude Code wrote essentially all of the code. I designed the game, directed Claude, shot at the monitor, and ruled on what was acceptable, but I did not write any code by hand.

The hard part was never the game logic. It was getting an agent to develop and debug a program that measures the physical world, when the agent cannot see the monitor or pull the trigger.

1. **An in-process MCP server, so Claude could ask the running app what it measured.** It exposes the latest camera frame, the detector’s arithmetic for every candidate blob including the ones it rejected and why, what is painted in both screen and camera coordinates, and key and click injection that does not steal focus from the full screen window. It returns measurements, never the app’s own conclusions, because the app’s conclusions were repeatedly wrong: it once printed “100% of measured shots clear the bar” on the same line as a measurement showing its own test dot below its own threshold. This took diagnosis from hours to about ninety seconds. Worked example: a build kept saying “unplug the camera”. Through MCP, Claude read a calibration measured seconds earlier at 0.91 mm, re-ran the self check and got 12 of 12, and traced the fault to one artwork file missing from the bundle, which was throwing an exception inside the paint handler. Because that exception escaped before the “this frame is on the glass” signal fired, every wait-for-paint timed out and the camera got the blame. 1,722 tracebacks in a single launch, and none of it in the application log.
2. **Making the tests able to fail.** The whole-launch harness had been exiting after its first launch and reporting success for as long as it had existed, so an “8 of 8 cold starts” figure had never been measured at all. The fix was mutation testing: reintroduce each known fault on purpose and require its check to go red. The first run put seven faults back and three survived. The project now does this 326 times across six runners, and all 326 are caught.
3. **A gate on the word “ready”.** Single good runs are misleading here. One calibration measured 0.4 mm held-out error and the next measured 273 mm with no code change in between. Nothing counts as working until five consecutive clean cold starts.

The honest headline number is not a test result: an independent microphone channel counts trigger pulls separately from the camera, and **84 to 89% of heard trigger pulls become a detected shot**.

All twenty lessons learned are described in details in the repo.

**Free to try, two ways to set up**

Open source (Apache-2.0), source plus a Windows release: [https://github.com/vkorost/screenplink](https://github.com/vkorost/screenplink)

1. Download the release, unzip it, and run *Play.cmd.*
2. Or open the unzipped folder in Claude Code and ask it to set up the app. The release ships a CLAUDE md and an .mcp.json for exactly that job: Claude runs the camera check, picks the camera actually pointed at your monitor, starts the app, and tells you either to shoot or which physical thing to fix.

*What you need*: **Windows**, a **matte** (anti-glare) monitor, a **webcam** that does **60 fps**, and a **red laser trainer**. Glossy screens and TVs do not work. 


```



document fetched 2026-10-26 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xeqtp1/i_built_a_laser_dryfire_shooting_game_with_claude
