Use when Codex is building or iterating on a web game (HTML/JS) and needs a reliable development + testing loop: implement small changes, run a Playwright-based test script with short input bursts and intentional pauses, inspect screenshots/text, and review console errors with render_game_to_text.
Build games in small steps and validate every change. Treat each iteration as: implement → act → pause → observe → adjust.
export SKILLS_ROOT="${LOBSTERAI_SKILLS_ROOT:-${SKILLS_ROOT:-$HOME/Library/Application Support/LobsterAI/SKILLs}}"
export WEB_GAME_CLIENT="$SKILLS_ROOT/develop-web-game/scripts/web_game_playwright_client.js"
export WEB_GAME_ACTIONS="$SKILLS_ROOT/develop-web-game/references/action_payloads.json"
Installed skills resolve from $LOBSTERAI_SKILLS_ROOT / $SKILLS_ROOT (production default: app userData/SKILLs, macOS usually ~/Library/Application Support/LobsterAI/SKILLs).
window.render_game_to_textwindow.advanceTime(ms). Strongly prefer a deterministic step hook so the Playwright script can advance frames reliably; without it, automated tests can be flaky.progress.md exists, read it first and confirm the original user prompt is recorded at the top (prefix with Original prompt:). Also note any TODOs and suggestions left by the previous agent. If missing, create it and write Original prompt: <prompt> at the top before appending updates.playwright is available (local dependency or global install). If unsure, check npx first.$WEB_GAME_CLIENT after each meaningful change; do not invent a new client unless required.$WEB_GAME_ACTIONS to avoid guessing keys.render_game_to_text reflects the same state shown on screen. If anything is off, fix and rerun.
Examples of important interactions: move, jump, shoot/attack, interact/use, select/confirm/cancel in menus, pause/resume, restart, and any special abilities or puzzle actions defined by the request. Multi-step examples: shooting an enemy should reduce its health; when health reaches 0 it should disappear and update the score; collecting a key should unlock a door and allow level progression.Example command (actions required):
node "$WEB_GAME_CLIENT" --url http://localhost:5173 --actions-file "$WEB_GAME_ACTIONS" --click-selector "#start-btn" --iterations 3 --pause-ms 250
Example actions (inline JSON):
{
"steps": [
{ "buttons": ["left_mouse_button"], "frames": 2, "mouse_x": 120, "mouse_y": 80 },
{ "buttons": [], "frames": 6 },
{ "buttons": ["right"], "frames": 8 },
{ "buttons": ["space"], "frames": 4 }
]
}
Test any new features added for the request and any areas your logic changes could affect. Identify issues, fix them, and re-run the tests to confirm they’re resolved.
Examples of things to test:
render_game_to_text JSON output.Expose a window.render_game_to_text function that returns a concise JSON string representing the current game state. The text should include enough information to play the game without visuals.
Minimal pattern:
function renderGameToText() {
const payload = {
mode: state.mode,
player: { x: state.player.x, y: state.player.y, r: state.player.r },
entities: state.entities.map((e) => ({ x: e.x, y: e.y, r: e.r })),
score: state.score,
};
return JSON.stringify(payload);
}
window.render_game_to_text = renderGameToText;
Keep the payload succinct and biased toward on-screen/interactive elements. Prefer current, visible entities over full history. Include a clear coordinate system note (origin and axis directions), and encode all player-relevant state: player position/velocity, active obstacles/enemies, collectibles, timers/cooldowns, score, and any mode/state flags needed to make correct decisions. Avoid large histories; only include what's currently relevant and visible.
Provide a deterministic time-stepping hook so the Playwright client can advance the game in controlled increments. Expose window.advanceTime(ms) (or a thin wrapper that forwards to your game update loop) and have the game loop use it when present.
The Playwright test script uses this hook to step frames deterministically during automated testing.
Minimal pattern:
window.advanceTime = (ms) => {
const steps = Math.max(1, Math.round(ms / (1000 / 60)));
for (let i = 0; i < steps; i++) update(1 / 60);
render();
};
f) to toggle fullscreen on/off.Esc to exit fullscreen.Create a progress.md file if it doesn't exist, and append TODOs, notes, gotchas, and loose ends as you go so another agent can pick up seamlessly.
If a progress.md file already exists, read it first, including the original user prompt at the top (you may be continuing another agent's work). Do not overwrite the original prompt; preserve it.
Update progress.md after each meaningful chunk of work (feature added, bug found, test run, or decision made).
At the end of your work, leave TODOs and suggestions for the next agent in progress.md.
playwright dependency if the project already has it.npx:
command -v npx >/dev/null 2>&1
npx is missing, install Node/npm and then install Playwright globally:
npm install -g @playwright/mcp@latest
@playwright/test unless explicitly asked; stick to the client script.$WEB_GAME_CLIENT (installed default: $SKILLS_ROOT/develop-web-game/scripts/web_game_playwright_client.js) — Playwright-based action loop with virtual-time stepping, screenshot capture, and console error buffering. You must pass an action burst via --actions-file, --actions-json, or --click.$WEB_GAME_ACTIONS (installed default: $SKILLS_ROOT/develop-web-game/references/action_payloads.json) — example action payloads (keyboard + mouse, per-frame capture). Use these to build your burst.