AI assistant

The assistant is an editor panel where you ask for a change in your own words: "add three coins", "change the colors to sunset". It builds the change, checks it and shows it to you before touching your game. You decide whether it is applied.

This page also says what the assistant does not do well. Read all of it before trusting it with a game.

Three ways to use it#

The first time you open the Assistant panel, the editor asks you to choose:

Choose how to use it opens Your AI assistant. The same thing is always in Settings.

What your installation includes#

The assistant with a model needs the Green Games AI layer, a program that goes with the editor. Not every package includes it:

Up to 0.2.0 the macOS package came without the AI layer; since 0.3.0 it includes it. Without the layer, the editor works in full and the assistant works in basic mode, and says so. The Your AI assistant screen tells you what it found:

The basic assistant#

Pick Basic assistant, no model and press Use this option.

It recognizes requests like these, which the assistant itself reminds you of when it does not understand something: "add a bridge that goes up and down", "make a lava level", "build me a block dragon character", "change the colors to sunset", "make it harder", "add 5 coins" or "place a tree".

If you also have the AI layer connected, it solves exact requests with its own operations, without a model: "add three coins", "put two enemies", "make it easier", "change the colors to ice", "put a tree and a house next to the start". They are fixed operations: they always do the same thing and are validated and tested with the bot.

What it does not understand, it says:

It works the same in English as in Spanish: in our session it understood "change the colors to sunset" and "add three coins", and answered in English.

Asking for a change and reviewing it#

Type the request at the bottom, in Ask the assistant for a change…, and press Enter. The assistant answers with a proposal:

A proposal has:

When applied, the change is a single step in the history: it is undone with Ctrl+Z or with Undo this change in the conversation. The conversation is saved with the project.

The Checked sentence says what was really reviewed. For a request that adds pieces it reads like this: "Checked: the game is still valid; the test bot finishes it; the new or moved pieces do not sit inside solid pieces that were already there." If the assistant had to move a piece because it sat inside another one, or if something could not be checked, the answer says that too.

The label can be one of five:

Up to 0.2.0 overlapping pieces were not checked. Proposals saved in a conversation from that version are shown with the label that matches what was checked back then, never as "verified".

Use Apply anyway only if you know what you are going to fix afterwards: with a proposal that failed verification, the game may end up impossible to finish.

Using a model on your computer#

You need the AI layer (see the table above) and two more things, which you provide:

1. Install Ollama. It is the free program that runs models on your computer. It does not ship with the editor: you install it separately, once, from its official site. The editor looks for it at http://127.0.0.1:11434 and, if it does not find it, shows the steps.
2. Go back to Your AI assistant and press Check again. The Ollama line must state the version and that it is running.
3. Under Model on your PC, press Download on the model you want. The download shows its progress. It never starts by itself.
4. Select the installed model and press Use this option.

These are the models editor 0.4.0 offers, with what the editor itself says about each:

The recommended one is qwen2.5-coder:14b. If your computer has no graphics card with that much memory, the editor tells you ("Does not fit on this computer") and suggests a smaller one, which gets a lot less right.

How often it really gets it right#

These are the measurements that exist. There are no others.

On the benchmark. Green Games measures the assistant with 41 fixed requests, each with a check that the requested change happened. With qwen2.5-coder:14b, on a computer with a high-end dedicated graphics card:

By type of request: creating a game, 6 of 9; editing an existing one, 8 of 13; fixing a problem, 5 of 5; building a piece or a character, 6 of 6; rejecting what does not belong, 6 of 8.

On free-form requests, written by hand. Here the results are uneven. Seven requests were tried with the same model inside the editor, twice:

• They ended in a valid change (verified and with the bot winning) five out of seven the first time and three out of seven the second.
• Two requests failed both times. Two passed once and failed the other time.
• Valid was not the same as correct. Of the five valid proposals of the first run, read one by one, only one did what was asked; another did half of it; and three were changes that worked but were not the request. The three of the second run did do what was asked.

In our session for this guide. With editor 0.4.0, on the Collection template, with qwen2.5-coder:14b:

We asked "add a purple decorative platform near the goal". It took 51 seconds (it was the first request: that includes loading the model). The proposal arrived with the labels "verified" and "bot finishes it":

But that template has no goal: you win by collecting all the coins. The model did not say so. It put a 3 by 3 decorative platform, with no physics, inside the purple stage that already existed: only one edge sticks out. The answer says the new pieces do not sit inside solid pieces that were already there, and the check did not flag it, because a piece with no physics collides with nothing:

The second request was more concrete: "put a stone wall 6 wide between the start and the goal, leaving a gap on the right". It took 16 seconds and put a single stone wall 6 wide, 2 high and 1 deep, over a lava zone. The overlapping check raised it 0.2 because it sat inside another piece, and the answer says so. The width and the material are the ones requested; there is no gap on the right: the wall can be walked around on both sides.

Two requests, two "verified" proposals, and neither did exactly what was asked. It is a small sample, but it matches the measurements above. The model's answers vary from one run to the next: this guide's Spanish session got different pieces for the same requests.

How to ask so it comes out better#

• One change per request. "Add three coins" and then "paint the wall red", not both things together.
• Be concrete. Say what, how much and where: "a stone wall 6 wide in front of the start". Use the names of the pieces you see in the Outline.
• Ask for what exists. If your game has no goal, do not ask for something "near the goal".
• Prefer exact requests when they fit: a fixed operation solves them, in seconds and always the same way.
• Look at Before and After in the scene, and read What changes. If it is not what you wanted, Discard costs nothing.
• Test after applying. Play for a moment. If something came out odd, Ctrl+Z.
• Do not ask it to touch the Rules. The editor writes the script's rules block; the model cannot change it and those requests fail. Change the rules in the Rules panel.

Limits#

• With the standalone Windows .exe there is no AI layer: only the basic assistant.
• In a Blocks project the assistant is turned off.
• There is no published cloud service.
• The model does not see the screen: it works on the game's files.
• Small models get a lot less right. With qwen2.5-coder:7b the benchmark gives 41%.
• With no graphics card, a model runs on the processor and is slow.
• The measurements are from a single computer and a single date. On another computer the times change.
• It does not make models, textures, drawings or sounds. If you ask it to, it answers that it does not generate them and says where the asset workshop and the importer are. It does not receive the list of what is in the workshop either.
• In a project with models, textures or sounds, the assistant keeps them as they are: it cannot write or delete them.

Next#

• No-code rules: what the assistant must not touch is changed here.
• Test bot: the test that backs the "bot finishes it" mark.
• What is missing in editor 0.4.0