# How to playtest Roblox games with AI

Writing a feature is half the work. The other half is proving it behaves. Playtesting is slow and repetitive: press Play, walk to the thing, click it, check Output, stop, edit, repeat. An agent that can drive Play mode itself takes most of that loop off your hands.

## What AI playtesting means here

When Kiwari is connected to Studio, it can do the same things you do during a test:

- Start and stop Play mode.
- Run a small check against the live running game, on the server or the client.
- Press keys, move the mouse, and walk the character to a spot.
- Read the console output, including prints and errors.
- Take a screenshot of the viewport and look at it.

Nothing from a Play session is saved into your place. When Play ends, the place goes back to exactly what it was.

## Why this matters

Without a test, "done" really means "it compiled". With one, the agent can find out that a button does nothing, a door never opens, or a script errors only on the client. It then fixes the problem and checks again, so you spend your time on the interesting decisions instead of the clicking.

## A good test prompt

Be specific about what success looks like:

> Run the game and check that touching the red coin adds 1 to the coin counter and removes the coin. Report what you see in Output.

Vague prompts such as "test everything" lead to vague checks. A concrete expectation lets the agent decide whether the result matches.

## Types of checks

### Console checks

The simplest and cheapest. Ask the agent to run the game, perform an action, and report errors and prints. This catches a large share of script bugs.

### Behavior checks

For game logic, the agent can inject a short snippet into the running game. For example, it can read a player's leaderstats after an action and confirm the number changed.

### Input checks

For things that depend on a human, such as pressing a key to jump or clicking a shop button, the agent can send keyboard and mouse input and move the character toward a target.

### Visual checks

A screenshot answers questions the console cannot: is the UI visible, is the model placed on the floor, is anything overlapping. The agent sees the image and can describe or judge it.

## A test-fix-test session

1. Ask for a feature.
2. Approve the diff.
3. Say: "Run the game and confirm it works. If it fails, fix it and try again."
4. Read the summary of what was tried and what the console showed.

If the agent changed the place and did not check afterwards, Kiwari asks it to either verify or say plainly that the change is unverified. That keeps unchecked work from being presented as finished.

## Limits to keep in mind

- **Studio must be open.** The agent plays inside your running Studio, not a separate server.
- **It is not multiplayer.** A single Play session tests one player. Networked edge cases still need real testers.
- **Feel is yours to judge.** An agent can confirm that a jump happens. It cannot tell you that it feels good.
- **Time boxes.** Long tests are capped, so keep each check focused.

## Review before it runs

Starting Play mode and running code are actions that wait for your approval by default, just like edits. You can allow them in Settings, Permissions once you trust a workflow.

## Tips

- Test right after each feature, while the change is small.
- Ask for the console output in the final summary so you can see what was checked.
- When a fix fails twice, ask for a screenshot. It often reveals a placement problem the logs hide.

[Start for free](https://kiwari.dev/download) and let the next feature test itself.

Related: [How to fix Roblox scripts with AI](https://kiwari.dev/blog/how-to-fix-roblox-scripts-with-ai) · [How to get started with Kiwari](https://kiwari.dev/blog/how-to-get-started-with-kiwari)