How to use the Test Playable Web Games agent skill

What the skill makes the agent do

Without a skill, an agent asked to test a game tends to load it once and say whether it looks fine. This skill changes that. It tells the agent to combine deterministic tests with short real playthroughs, and to treat a passing build as only one piece of evidence rather than proof that the game works.

The agent first maps the player journey: launch, new game, controls, core action, enemy encounter, reward, inventory and progression, save and continue, loss and retry, pause, settings, and completion. It also adds device variants for desktop keyboard and mouse, touch portrait, touch landscape, and reduced motion.

• Jumps straight to specific states (combat phase, boss phase, low health, save migration, error states) through seedable fixtures, debug routes or query parameters instead of grinding through the campaign.
• Confirms that the screen responds after each meaningful action: control response, target state, damage or resource state, UI update, and navigation.
• Inspects console warnings and errors, and samples performance only on a representative live scene.
• Records reproduction steps, expected versus actual result, severity, device and viewport, and the shortest useful proof.
• Separates new regressions from pre-existing baseline issues.
• Closes idle browser pages, dev servers and benchmarks after testing, while leaving active task resources alone.

Install the skill from the ZIP

Download the test-playable-web-games ZIP from this page. The skill package contains two files: SKILL.md, which holds the instructions, and agents/openai.yaml. Unzip it so the folder named test-playable-web-games sits inside the skills directory you want to use.

For a personal install, unzip into ~/.claude/skills/. To scope it to one project, unzip into that project's .claude/skills/ folder instead. You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/test-playable-web-games/
• Project only: <your-project>/.claude/skills/test-playable-web-games/
• Claude apps: Settings → Capabilities → Skills, then upload the skill

Check what you need before you start

The skill needs two things. First, a browser surface the agent can control. The skill says to use the repository-approved browser tooling, so your project should already have a browser tool set up for the agent. Second, a running dev or staging build of the game to test against. The skill does not start your game for you based on its own text, so have the build running before you ask.

• A browser surface the agent can control, using the repository-approved browser tooling
• A running dev or staging build of the game
• Deterministic test hooks help a lot: seedable fixtures, debug routes or query parameters for the states you want to reach

Trigger the skill with natural requests

You do not need special syntax. Describe the testing job in plain words that match the skill's description, such as gameplay QA, regression testing, controls, save flows, console checks, responsive or mobile testing, performance smoke tests, or release verification. Name the game's address or the area you care about so the agent knows what to test.

Be specific about scope. A request for a full pre-release pass produces a different run than a request to check one save flow.

• "Run a pre-release QA pass on the game running at localhost, covering keyboard, touch portrait and reduced motion."
• "Test the save and continue flow, including what happens after a loss and retry."
• "Check the console for errors while playing through combat, and sample performance on the arena scene."
• "Write up the bugs you find with steps to reproduce, device and viewport, and whether each one is new or was already there before."

Read the output: bug reports and evidence

The skill asks the agent to report actionable evidence. Each finding should include reproduction steps, expected versus actual result, severity, device and viewport, and the shortest useful proof. Findings are split into new regressions and pre-existing baseline issues, so you can decide what blocks a release.

Review the evidence rather than trusting the summary. Check that the steps are concrete enough for another person to repeat.

• Check each finding has steps, expected versus actual result, severity and device/viewport
• Confirm the regression list and the baseline list are separate
• Ask the agent for a shorter proof if the evidence is hard to follow

Safety notes: scripts and network

This skill does not run scripts and does not use the network, according to its listed details. Its files are SKILL.md and agents/openai.yaml. Because it works through a browser the agent controls and a build you run, the main risks are the same as for any browser testing: the agent may open pages and interact with your running game. Use a dev or staging build, not production accounts or real player data.

• Test against a dev or staging build, not live player accounts
• The skill itself runs no scripts and makes no network calls, per its listing
• Review what the agent does with the browser tooling your repository approves