AI GamingFinalistGPT & Claude

PlayerOne: an AI playtester that reports bugs with proof

Indie and small game studios with 1 to 15 people who build PC games in Godot, Unity or Unreal and have no QA team. They test by hand, and nobody has time to replay the same level a hundred times after every build. So soft-locks, crashes and broken counters slip through to players. The first bug report is often a one-star review with no steps to reproduce. Hiring testers is slow and expensive. Letting a big AI play the game is too costly to run on every build, because it reads a full screenshot for every single move. PlayerOne gives these teams a tester that plays the game on their own machine with a free local AI. It reads the engine's own error console and writes a bug report with the error, the steps to reproduce, the frames before the bug and a short clip. A big AI like Claude, Codex or Gemini only checks small summaries, so it costs about 96% fewer tokens than letting it play.

Open the demo Video GitHub

Preliminary score

81.5/ 100Strongaverage of 2 judges

Value for the user22.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI21.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing17.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility13.0 / 15
Data requirements, running costs and a clear next step.
Originality7.5 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 70.0 / 100 on stage.

What the judges wrote

Claude Opus 5.5Preliminary82/100
Strengths
Your target user is sharp: small PC studios with no QA team, who learn about bugs from one-star reviews with no repro steps. Your benchmark is exemplary for a hackathon. You planted five bugs, added a clean build to count false alarms, used a random-keys baseline, ran two rounds with logs, frames and clips in the repo, and honestly fixed and documented failures such as the capture freeze and the 13 wall-bump false alarms. The 'cheap hands, smart judge' split is a clear design, and measuring 2,975 vs 70,026 tokens per minute gives the cost claim real grounding. Each report ships with a stack trace, recorded-input repro steps, a contact sheet and a clip, which is a concrete, useful output.
To improve
By your own numbers, the non-AI explorer bot and rule-based detection catch 4 of 5 bugs, while the local 3B model caught only the soft-lock. Show the big-model judge actually catching something the rules cannot, such as the HUD coin-counter bug, to prove AI is essential and not optional. Run several trials per cell and test on at least one third-party game, because one run per bug on self-planted bugs is a thin sample. Also re-benchmark the new control detection and mouse play, since they are not covered by the current numbers.
GPT-6 AstraPreliminary81/100
Strengths
Your workflow addresses a concrete small-studio problem by packaging engine errors, recorded inputs, frames and clips into actionable bug reports. Your before-and-after benchmark compares three drivers, documents missed bugs, and explains fixes for capture hangs and wall-related false alarms. Your console-first detection and compact AI digests provide a credible path to affordable testing after each build.
To improve
Your next benchmark should repeat runs on third-party games and include timed human QA, measuring coverage and false-positive rates beyond one run per test condition. Your AI contribution needs a clear ablation showing whether the cloud judge improves incident precision, semantic-bug detection or report usefulness over the stronger non-AI explorer. Your cost model should include output tokens and local hardware requirements, while compatibility claims should distinguish tested Windows/Godot support from untested builds and planned engine probes.

The team

Nijat HatamliLead
developer
Aliyyiakbar Shirinli
data
Aylan Macabalitao
designer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Qwen2.5-VL 3B running locally through Ollama as the free player. Claude, OpenAI and Gemini APIs as optional players and judges, using the user's own key. Claude was also used as a coding assistant while building.
Data: No training data. Our own Godot test game built during the hackathon with five bugs planted on purpose. All test results come from our own runs and are in the repo under desktop/bench.
Components: Rust, Tauri 2, xcap, enigo, reqwest, image, clap, regex, window-vibrancy, Vite, TypeScript, React, Tailwind, Framer Motion, Lenis, Godot 4, ffmpeg, Ollama, Supabase, Stripe test mode, Netlify, GitHub Actions. No public templates.
Built during the event: Confirmed

Quality testing

We built a small Godot game and planted five real bugs in it: a soft-lock, a script error, a crash, a full freeze and a coin counter that stops updating. We also made a clean build with no bugs to count false alarms. Every run was one minute of real play. We compared three players: random key presses as the baseline, our free explorer bot with no AI, and a local vision AI (Qwen2.5-VL 3B). All 18 runs, with a session log, frames and clips for each, are in the repo under desktop/bench. First round results Soft-lock: random 38s, bot 17s, AI 9.6s Script error: only the bot found it, in 22s Crash: only the bot found it, in 16s Freeze: PlayerOne froze together with the game in all three runs Coin counter: missed by everyone Clean build: 13 false alarms across all runs What broke and how we fixed it 1. PlayerOne froze when the game froze. Grabbing a frame waits for the game to draw, and a hung game never draws. We moved capture to a separate thread with a 1.5 second deadline. A window that stops drawing is now reported as a hang. 2. 13 false alarms. Our stuck detector fired every time the player simply walked into a wall. Now PlayerOne first pauses and tries every movement key once, like a human tester would. If anything moves, it was a dead end and nothing is filed. 3. The Windows "not responding" flag was wrong. It was raised on healthy game windows, so we stopped trusting it. Second round results Soft-lock: random 34s, bot 15s, AI 39s Script error: bot 12.5s Crash: bot 18s Freeze: bot 20s Coin counter: still missed False alarms: 0 in every run The bot now catches 4 of 5 bugs with zero false alarms. Random keys catch 1. Failures we have not fixed The coin counter bug gives no error and no freeze, so only an AI that understands the screen can catch it, and the 3B model did not. The local AI is also slow. It makes 25 to 45 moves a minute against 77 for the bot, so it rarely reaches later bugs in one minute. The cheap bot is the better driver, and the AI is better used for judging. When we used the 3B model as the judge it wrote weak repro steps like "press a key", so PlayerOne now uses the recorded inputs instead. The new control detection and mouse play were added after these tests and are not covered by these numbers. Comparison with how it is done today A human tester replays the level and writes the ticket by hand. PlayerOne found the crash, the freeze and the script error in 12 to 20 seconds. Each report came with the exact stack trace (main.gd line 148), repro steps, a 36 KB picture of the frames before the bug and a 14 KB clip. Letting a big AI play through computer use costs about 67,000 to 85,000 tokens per minute of play, because it reads a full screenshot for every move. With PlayerOne the big AI only judges. In our test it used 2,975 tokens for the same minute instead of 70,026. That is about 96% fewer. Everything it received for that minute was 45 KB.

Worth 20 of the 100 points.

Running it

Download the app from the website or from github.com/nijathatamli/AI-hakaton/releases/latest. There are builds for Windows x64, x86 and ARM64, macOS and Linux. The builds are not signed yet. On Windows click More info and then Run anyway. On macOS right click the app and choose Open. To play for free with a local AI, install Ollama and run ollama pull qwen2.5vl:3b. Without it a free bot plays instead. To use a big AI, paste a Claude, OpenAI or Gemini key in the Connect page. Open your game, open PlayerOne, click the game window, check the controls it suggests and press Start. Press F8 or the red Stop button to stop. Bug reports appear in the Reports tab with the error, the repro steps, the frames and a clip. To see it catch planted bugs, install Godot 4 and run this in the desktop folder: playerone run --launch "godot --path testgame -- --bugs=crash" --window Cavern --engine godot --preset platformer --minutes 1 The free Free plan needs no account. Sign in from the Account page to use your own AI keys. Known limitations: only the Windows build was tested on real hardware. The others come from automated builds. The small local AI plays 2D games well but struggles in 3D, so a big AI gives much better results there. PlayerOne sends real key presses to the game window, so do not type while it runs. On Linux it needs X11. On macOS it asks for Screen Recording and Accessibility permission the first time.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.