Studies
This is a copy of the example study “Study 3: Intrasender role conflict” for you to edit. You can change every setting. Nothing is stored until you click Save changes. The first save creates the study with a new ID and password, and after the first save you can test the API key and use the AI Assistant.
You are working without an account. Creating an account lets you manage your studies more easily: all your studies appear in one list on the studies page and open from any device without typing the study ID and password.

Study settings

Set up your study in four steps. You can switch between the steps at any time. Changes apply to new chats as soon as you save the changes.

Step 1 · Study setup

Start with the basics of your study: a name, background information for the AI Assistant, and what participants see before and during the chat.

Study

Only you and your collaborators see the study name. Participants never see the study name.

The study ID identifies this study in all links and in the exported data. MAESTRO creates the study ID automatically, and the study ID cannot be changed.

Study context

Describe your study in a few sentences. The AI Assistant in Step 3 reads the study context when helping you write the reference script and the conductor instructions. The conductor (the model that adapts the conversation during the chat) and the participants do not see the study context.

The more the AI Assistant knows about your study, the better the texts from the AI Assistant fit your research. Useful content is a one- to three-sentence summary of the research idea, the concepts and definitions you work with, important literature, and preferences about the conversation, for example that all teammates should write in Spanish. The AI Assistant takes such preferences into account when writing or changing texts for you.

The conductor, which rewrites the conversation during the chat, does not see the study context. Anything the teammates must do during the chat, such as writing in Spanish, therefore has to be part of the conductor instructions or of the reference script in Step 3.

Session flow

Here you choose what participants see before and during the chat.

What participants see before the chat

Participants click “I'm ready” and see “Finding teammates…” for a moment, as if they were being matched with other people.

A wait of 10 to 30 seconds makes the matching feel real without trying the participants' patience. The waiting time sets how long “Finding teammates…” is shown before the chat opens.

Participants see a short description of the task above the “I'm ready” button. The description is only shown when the waiting screen is switched on.

What participants see during the chat

Participants see how much time is left. The conductor is told the remaining time as well, so your instructions can refer to the remaining time, for example to wrap up near the end.

When the timer reaches zero, the chat stays open and participants can keep writing. If the conversation should end at that point, you can start a new segment in the reference script (Step 3) shortly before the end, for example at 9:30 with the prompt “The team says goodbye in one or two short lines and then stops writing.”

Participants see a formatted text in a box between the timer and the chat. The box stays visible for the whole session, so participants can look up what they need to do at any time.

Instructions for participants

You can select text and use the buttons to make the text bold, italic or underlined, to turn lines into a list, or to change the size and color of the text. The box in the chat looks like this editor and has the same width on a computer screen. Text that you paste from another program keeps its lists and its basic formatting.

Bold, italic and underlined also have keyboard shortcuts: ⌘B, ⌘I and ⌘U on a Mac, and Ctrl+B, Ctrl+I and Ctrl+U on Windows. With text selected, a shortcut formats that text. Without a selection, the shortcut switches the formatting on or off for the text you type next.

Short texts work best. The box takes up at most about a third of the window, so participants have to scroll inside the box when the text is longer. If you leave the text empty, no box is shown.

Everything the AI teammates need to know about the task belongs in the conductor instructions in Step 3, because the AI teammates do not read this text.

Participants see this text as it looks here. You can open the test chat after saving to check the text.

Step 2 · AI model

Choose the language model that plays your AI teammates. This model works as the conductor: each time the participant sends a message, the conductor rewrites the rest of the conversation. The model's speed decides how quickly the team reacts, and the model's price decides what each conversation costs.

Provider and model

Pick the provider (the company that runs the model) and the model. Step 4 shows what your study costs with the model you pick.

OpenAI and Anthropic are the companies that make the models. You need an API key from the provider you choose (a key from your account there that lets this tool use the model), and that provider bills your account for every request. You can choose Custom if your university or company offers OpenAI models through its own gateway (a server that passes requests on to the model).
Smaller models answer faster and cost less. Larger models follow long and detailed instructions more reliably, but participants wait longer for the first reply. Try a few test conversations before you collect data, and keep the same model for the whole study so that all participants get a comparable experience.
Many models can think before they write. Thinking can improve the replies, but every second of thinking is a second in which the chat stays silent, and the thinking is billed like written text (as output tokens, the billing units for text the model produces). The fastest setting is a good default; a slower setting is only worth trying if the replies are clearly poor.

This sets how long the model may think before the team reacts to a message. The fastest setting is recommended.

API key

The key connects this tool to your account with the provider. The key stays on the server with this study: participants never see the key, and this page shows only the last four characters. The Test button checks that the key and the model work.

OpenAI keys start with sk- and are created at platform.openai.com under API keys. Anthropic keys start with sk-ant- and are created at console.anthropic.com under API keys. The account also needs credit (a prepaid balance or an active payment method), otherwise requests fail even with a valid key.

The test sends one very short request to the selected model. The test costs a tiny fraction of a cent and uses the key in the field, even if you have not saved the key yet.

The time and the result of the last test are stored with the study right away, without pressing Save. A test that is more than a week old is shown in yellow here and in the readiness check of Step 4, as a reminder to test again before you collect data.

This decides in which of OpenAI's two request formats messages are sent to the model. Automatic works for OpenAI itself. Some gateways send the whole answer at once instead of word by word when the older format, Chat Completions, is used, which makes the team react late. The newer format, the Responses API, avoids the delay, and the university gateway currently needs the Responses API.

Step 3 · Conversation

In this step you set up the conversation your participant will have. The participant joins a group chat with several AI actors, LLM-based teammates with their own personalities and communication styles, to work on a task in a situation you describe. The way the AI actors treat the participant creates the social dynamic of your study, and the dynamic is where your conditions differ. The conversation follows a reference script that you write in advance: the actors' messages with their timing, which the AI actors play line by line. A fixed script sounds robotic the moment the participant says something unexpected, so an orchestrating language model, the conductor, sits above the actors. Whenever the participant writes a message, the conductor rewrites the messages that are still to come, so that the actors respond to the participant while the situation, the personalities and the dynamic stay as you defined. The cards below guide you through these pieces one by one, and the conductor reads the cards as one prompt.

Situation: What is the task during the conversation, and what is the background?

Suggested instructions

AI actors: Who are the AI actors, and how do the actors behave?

Instructions for all AI actors
Suggested instructions

Dynamics: Which social dynamic should the AI actors create?

Suggested instructions from the three published studies

Conductor: How should the conductor run the chat?

The conductor is the language model that keeps the conversation coherent. Whenever the participant writes a message, the conductor rewrites the messages the AI actors have not sent yet, after reading that message, the conversation so far, the reference script and the texts on this page. The actors then respond to the participant while the situation, the personalities and the dynamic stay as you defined. The conductor stays invisible to the participant and never writes in the chat itself. The boxes below tell the conductor what to achieve and what to keep in mind when the conductor rewrites the conversation.

Purpose: What should the conductor achieve?
Suggested instructions
Rewrite rules: What should the conductor keep in mind when it rewrites the conversation?
Suggested instructions
Output format: How should the conductor format the output?
Expert
Suggested instructions

Reference script: What should a typical conversation look like?

Each row is one message. The number in seconds is the pause after the previous message, and the percentage says for how much of that pause the typing indicator is shown. The ghost button makes a teammate type for a moment and stop without sending anything. You can drag rows to change their order, or switch to JSON (the script as structured text) to paste a complete script.

The AI Assistant can write a new script or change the existing script for you, including the segments. The participant's messages are not part of the reference script.

A segment starts at a fixed time, counted from the start of the chat after the waiting screen. At that moment the conversation is cut: whatever the team had planned is dropped, and the conductor continues with the segment's prompt (a short text that only says what changes, in 1 to 3 sentences). The conductor instructions below still apply.

Until the participant writes for the first time, the reference script simply plays as written. In each segment the first message's pause counts from the segment's start, so changes in one segment do not move the next segment. Messages that would come after the next segment's start are cut and shown crossed out.

To divide the conversation, you can use the scissors on the message where a new segment should begin, or add a segment at the end. At each segment start the conductor rewrites the conversation once more, which adds a little to the costs of every conversation.

Full conductor prompt

Whenever the human participant sends a message, the conductor receives the texts above together with extra information from the running chat, such as the conversation so far, the messages that were still planned and the message the participant just sent, and adapts the script in real time. Click "Show the full prompt" to see and audit the complete prompt the conductor receives.

Step 4 · Launch

Check that everything is ready, see what the study will cost, and share the link with your participants. You can pause the study here at any time.

Readiness check

These points are checked against the settings on this page, including unsaved changes.

    Study status

    Pausing the study stops new participants from starting, for example between two waves or once you have enough data. When the study is active, everyone with the participant link can take part.

    When the study is paused, the participant link shows a message that the study is not active, instead of the chat. Participants who are already in a chat can finish the chat. Test links keep working, so you can try changes before you continue the study.

    Pausing and continuing take effect immediately. You do not need to press Save for pausing or continuing.

    Costs

    The provider of the AI model charges for every conversation. Enter the number of participants and messages you expect, and the estimate shows what the study will cost.

    The estimate uses the conductor instructions and the reference script from Step 3 as they are now, including unsaved changes, and the model from Step 2. The count of tokens (the billing units of text) is usually within 5 percent of the provider's own count.

    Early in a conversation the conductor rewrites almost the whole reference script, so the output is large. Later there is less left to write, but the chat so far is longer. Providers bill the text the model writes (output tokens) at a higher rate than the text it reads (input tokens).

    For OpenAI models the costs include the prompt cache: OpenAI stores the beginning of each prompt (the text sent to the model) for a while. From the second message of a conversation on, the part that stays the same (mainly your conductor instructions) is read from the cache at a much lower price. The rest of the prompt is written to the cache, and with the newest models writing to the cache costs 1.25 times the normal input price. Anthropic models are billed without a cache.

    “Spent so far” counts every request this study has sent to a model: the conductor's rewrites in all conversations, including your test conversations, and the AI Assistant in Step 3.

    Model

    Estimated cost per participant
    Estimated cost of the study
    Spent so far

    Collected data

    Loading the number of conversations.

    View data
    All changes saved