The AI designer
The AI designer is a chat with a language model that designs missions with you. Tell it what you want, in plain words, and it builds it on the map while you watch. It works through the studio's own editing, so everything it does is an ordinary change: you can undo it, it shows in the change log, it saves, and it goes into the game only when you apply.
Open it with the AI designer tab on the right edge, beside Selection.
It needs an API key
The AI designer uses your own account with OpenAI (or another compatible provider): see Settings below. Everything else in the studio works without it.
Ideas and Build
The panel has two modes.
- Ideas: it looks at the level (its named places, buildings, guards, alarms and walkways) and answers with mission designs: a pitch, objectives, enemies and where they stand, security, time and weather, and difficulty. Nothing changes. Each design has a Build this button.
- Build: say what the mission is, and it builds it. For example: "A stealth mission at the radar dome: two snipers on the approach, a camera on the dome, reach the dome, light snow, 8 minutes." It looks at the places it will use, says in a few lines what it will build, then builds it.
Each step is a row in the chat, with Show to find it on the map, and appears on the map as it happens. Follow on the map moves the map along with it. At the end it runs the same check list as Apply, and fixes what it finds.
Then carry on: "make it harder", "move the snipers so they cover the road". It reads the mission as it is, your own changes included, and what you have selected on the map is part of what it knows: "put a sniper up there".
- Stop ends a request at once.
- Undo these changes, under a request, puts the mission back as it was before it. Ctrl+Z brings the changes back.
- On one of the original missions it can give ideas, but asks you to make a mission of your own before it builds.
Showing it what you mean
The buttons above the message box:
- Tag a place: drag an area on the map, or click a thing or a spot. Tags are lettered A, B, C, drawn on the map and sent with your message, so "a sniper nest in A facing the road" means exactly that.
- Sketch on the map: draw on the map in red, yellow, cyan or white, and say in your message what each colour means. "A fence along the red lines, a camera at the north-east corner" builds just that.
- Attach the map: the map as it is on screen, with a grid in metres.
- Attach a picture: a file, a pasted image or one dropped on the panel. A sketch on paper, a plan, a screenshot of another game's map.
What it can do
It can look at the mission and the level, find named places, look around a point, and check the mission. It can place and change guards and their patrols, structures, fence and wall runs, a fenced compound, pickups, cameras and alarm parts; move and remove things, yours or the level's; set objectives, events, the time limit and weather, the mission's name and the player start; lay new walkways and shape the ground; and move the map or open the 3D view to show you something. It keeps to the game's rules, as you would: guards need a walkway nearby, the map computer lists six objectives, and so on.
A few things worth asking for:
- A stealth check: "is there a quiet way in?" It finds the quietest way on foot from the start to each objective, says how much of it is watched and by whom, and suggests fixes, or makes them.
- Texts: the mission's name and description in the game's list, objective texts, map labels, and a short briefing shown at the start, in the game's terse style.
- New things for the inventory: "a sandbagged sniper nest I can keep", "a Spetnaz officer with a Desert Eagle". It makes a structure out of the game's own parts, or a character with a model, weapon and sight of its own, as a card in the chat with its 3D model. Place puts it on the map; Add to inventory keeps it for any mission. It cannot make new weapons or new 3D models: those come from the game.
Versions and campaigns
- Mission, Easy, normal and hard versions... makes easier and harder copies of a mission in your library: fewer or more guards, shorter or longer sight, more or less time, and so on. The guards the objectives and events name are kept.
- Mission, Make a campaign... puts missions in order under a name and a briefing, and exports them as one file. Importing it adds them in order.
The AI designer can do both for you too.
A draft without the AI
Mission, A draft from a description... reads a plain description with a few rules on your own computer, and nothing is sent anywhere. For example: "Two snipers near the radar dome, four Spetnaz guards at the barracks and two cameras by the control tower. Hack the terminal, grab the Dragunov at the watchtower, then escape by the north gate. Snow, a 15 minute time limit, don't raise the alarm." Read it lists what it understood; Add to the mission places it, as one step you can undo.
Chats
Each mission keeps its own conversations, as many as you like. The chat's name at the top of the panel opens the list: open one to carry on where it stopped, rename it, or delete it (click twice). + starts a new chat.
Settings
Open Settings (top right), AI designer:
- Provider: OpenAI, or OpenAI-compatible, for any server that speaks OpenAI's API: a model on your own computer through LM Studio, Ollama or vLLM, or a service such as OpenRouter.
- With OpenAI: your API key and the model. The key is encrypted for your Windows account and kept by the studio, never in its settings file, and the page never sees it again: it only shows the last four characters. Forget removes it. The models on your account are offered, newest first.
- With OpenAI-compatible: the model, the server's address (LM Studio's is
http://localhost:1234/v1), and its context window: how much the model can read at once, in tokens, as the server loaded it (LM Studio calls it Context Length). The studio fits every request into it: in a long chat the oldest parts are left out first, never the newest. An API key only if the server asks for one. - Thinking: how long it thinks before each step, from off to high. Pace: how long each step stays before the next, so you can watch it build. Steps per run: the most steps one request may take.
- Test saves, then sends one short request and says what came back. Save keeps the settings.
Light model, the next page of Settings, is a second model with the same choices: a quick, cheap or local one for the chores, naming chats, reading pictures, and looking things up for the main model. A chat can also run on it entirely: click the model's name in the panel. With OpenAI and no key of its own, it uses the main model's.
It uses your account
Requests go from the studio to your provider with your key, and your provider charges for them. A build of about twenty steps takes around 40 seconds and 100 to 170 thousand tokens, half of them usually from the cache.
Next: applying and playing.