Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Other skills

boot-game

Run your game and its services from the command line, take screenshot renders, and share the GPU safely with the user and other agent sessions. Use before running, rendering or timing anything in a game, or when the user says "run the game", "start the game", "take a screenshot of the scene", "render this level" or "boot-game".

off-grid-ai · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add off-grid-ai/skills --skill boot-game

About this skill


name: boot-game description: Run your game and its services from the command line, take screenshot renders, and share the GPU safely with the user and other agent sessions. Use before running, rendering or timing anything in a game, or when the user says "run the game", "start the game", "take a screenshot of the scene", "render this level" or "boot-game".

Start your game

Before the first run, find out how this project starts. Read the README, any run or launch scripts (*.sh, Makefile, package.json scripts, .vscode/launch.json), CI config and the engine project file. Write down what you find in the project's docs if it is missing. Ask the user if you can't find a way to run it.

  1. Native only. Run the engine build that matches the machine. On Apple Silicon, an agent's shell may run under Rosetta: check with uname -m and launch the engine as arch -arm64 <engine> ... if needed. An emulated engine is several times slower and every timing lies.
  2. The game's services. Many games have parts that run beside the engine: a local server, a simulation, a database, a dev web server. Find how they start, stop and report status (a script, a service manager entry, a docker compose file). Ask the user before you restart anything while they are playing. Once they say they aren't playing, restart freely.
  3. Renders from the command line. Find or add a screenshot mode: launch arguments that open a given scene, wait for an event, save a PNG and quit. Godot: user arguments after -- read with OS.get_cmdline_user_args(). Unity: -batchmode is headless, so use a normal player or editor run with -executeMethod and ScreenCapture. Unreal: -ExecCmds with HighResShot. Web: a headless browser such as Playwright. Prefer shots that time themselves to the event they show. A fixed wait is only the fallback. Document the flags in the project.
  4. One GPU run at a time, and the user's game comes first. If several sessions render on one machine, put every render through one queue script that holds a shared lock and waits while the user's own game is open. Never kill other sessions' render processes to make room. Run renders in the background, in sequence. Stop only your own renders, matched by their full arguments (pkill -f "<engine> <your args>"), never by the engine's process name alone: that kills the user's game too. A GPU timeout or driver stall usually means rerun it.
  5. Headless checks are not renders. Compile checks and headless probes don't use the GPU. A few may run at once (cap them, for example pgrep -fc "<engine> --headless" below 3). Bound every one so a probe that never quits can't hold anything up: use timeout, or on macOS without it perl -e 'alarm shift; exec @ARGV' 180 <cmd>, and give each probe its own in-game watchdog that exits with an error after a set time. A probe must quit on every path. Never queue a headless run behind a windowed render.
  6. Local services the game calls. If the game calls a local AI model, a game server or another app, find its address and how to start it. Note that the first call may be slow (a model loading). Ask the user to start anything only they can start.
  7. Mind the machine. Check uptime and the top processes before you time anything. Other builds and models running at the same time make numbers noisy.
  8. Assets in. Find the project's asset import rules (folder layout, import settings, retarget or rig setup, licence files) and follow them for every new asset. Write them down if they aren't written.
  9. On Linux or CI. If the project has a setup script for a fresh Linux box or CI, use it. A headless box usually needs a virtual screen and a software renderer for renders, which are slow; logic checks are not. Third-party art kept out of git must be copied in before the engine can open the project.