AI Prompt

Version: v1 (current)

The AI Prompt task presents participants with an interactive chat interface powered by a large language model (LLM). Researchers configure the AI's behavior, and all messages are recorded for analysis.

Overview

This task enables conversational data collection where participants interact with a configured AI assistant. The researcher defines:

• A system prompt that controls the AI's persona, tone, and behavior (invisible to participants)
• Optional RAG documents to give the AI specialized knowledge
• Safety filters and cost guardrails (turn limits, token budgets)

The complete conversation transcript (every participant message and AI response) is saved as task data, along with metadata such as response latency and token usage.

How It Works

1. Main Instructions

Before the chat begins, participants see your main instructions page explaining what they should do.

2. Conversation

Participants type messages and receive AI responses in a two-column chat interface:

• Left panel: the conversation area where messages appear in a scrolling chat window with participant and AI avatars, an optional typing indicator, and optional timestamps, turn counter, and Markdown rendering
• Right panel: a persistent instructions panel showing researcher-configured guidance, with the Finish button at the bottom

3. Completion

Participants click the Finish button in the right-hand panel when done. The full transcript is saved automatically.

Configuration Options

Provider & Model Settings

Option	Description	Default
AI Provider	Which AI provider to use	Google Gemini
Model	Model to use (options change based on provider)	gemini-2.5-flash
Temperature	Controls randomness (0 = deterministic, 2 = very creative)	0.7
Max Tokens	Maximum tokens per AI response	1024
Top-P	Nucleus sampling parameter (Gemini, OpenAI, Anthropic)	0.95
Top-K	Top-k sampling parameter (Gemini, Anthropic only)	40
Frequency Penalty	Penalizes repeated tokens (OpenAI only, -2.0 to 2.0)	0
Presence Penalty	Penalizes tokens that have appeared (OpenAI only, -2.0 to 2.0)	0

Available providers and models:

Google Gemini (default):

• gemini-3-pro-preview — Most capable preview model
• gemini-3-flash-preview — Fast preview model
• gemini-2.5-pro — Stable, most capable
• gemini-2.5-flash — Stable, fast, general-purpose (recommended)

OpenAI:

• gpt-4.1 — Most capable
• gpt-4.1-mini — Fast, cost-effective
• gpt-4.1-nano — Fastest, most affordable
• gpt-4o — Multimodal
• gpt-4o-mini — Fast multimodal

Anthropic:

• claude-sonnet-4-5-20250929 — Claude Sonnet 4.5
• claude-haiku-4-5-20251001 — Claude Haiku 4.5

The configuration panel adapts based on the selected provider — parameters that are not supported by a provider are automatically hidden.

note

Only Google Gemini is currently active in the backend. OpenAI and Anthropic support is planned for a future update.

Cost & Length Controls

Option	Description	Default
Max Turns	Maximum number of exchanges (0 = unlimited)	0
Total Token Budget	Overall cap on tokens for the conversation (0 = unlimited)	0
Max Message Length	Character limit for participant messages	2000
Context Overflow	Strategy when conversation exceeds context window	Summarize

Context overflow strategies:

• Summarize — Older turns are summarized to fit the context window (recommended)
• Truncate — Oldest turns are dropped
• None — Let the API handle truncation

Chat UI Options

Option	Description	Default
Initial AI Message	Optional greeting shown at the start	(empty)
Show Typing Indicator	Animated dots while AI generates a response	Enabled
Enable Markdown	Render AI responses with Markdown formatting	Enabled
Show Timestamps	Display time next to each message	Disabled
Show Turn Counter	Display "Turn X / Y" indicator	Disabled

System Prompt

The system prompt defines the AI's behavior. Participants never see it. Use it to:

• Set a persona ("You are a helpful tutor specializing in mathematics")
• Define conversation rules ("Always ask one follow-up question")
• Restrict topics ("Only discuss the provided documents")
• Control response style ("Keep answers under 3 sentences")

Safety Filters (Gemini only)

Gemini's built-in safety filters can be configured per category. This section is only visible when Google Gemini is selected as the provider.

Category	Options
Harassment	Block none / Block only high / Block medium and above / Block low and above
Hate speech	Block none / Block only high / Block medium and above / Block low and above
Sexually explicit	Block none / Block only high / Block medium and above / Block low and above
Dangerous content	Block none / Block only high / Block medium and above / Block low and above

Default is Block medium and above for all categories.

Instructions

Use the rich text editors to customize:

• Main Instructions: Explain the task before the chat begins (shown as a full-page pre-task screen)
• Hint Instructions: Quick reference available via the "?" button during the task
• Panel Instructions: Persistent instructions displayed in the right-hand panel throughout the conversation. Use this to remind participants of the task goal, what to discuss with the AI, and when to click Finish

Advanced Context Settings

These settings control the scaffolding text injected into the AI model's context. Customizing them prevents language bias when conducting studies in non-English languages.

Setting	Description	Default (EN)	Default (FR)
RAG Context Prefix	Text prepended to RAG documents in the system instruction	"Reference documents (use these to inform your responses):"	"Documents de référence (utilisez-les pour éclairer vos réponses) :"
Summarization Prompt	Instruction sent when older conversation turns need summarizing	"Summarize the following conversation concisely, preserving key topics, decisions, and context:"	"Résumez la conversation suivante de manière concise, en préservant les sujets clés, les décisions et le contexte :"
Summary Label	Label prepended to the generated summary	"[Summary of earlier conversation]:"	"[Résumé de la conversation précédente] :"

When the study's primary language is not English, update these strings to match the study language. This ensures all AI context is in the same language as the participant interaction.

Use Cases

Conversational Assessment

Evaluate comprehension through dialogue:

System prompt: "You are an interviewer assessing the participant's
understanding of climate change. Ask open-ended questions and
follow up on their responses. Cover causes, effects, and solutions."

Interactive Tutoring Research

Study learning through AI-guided instruction:

System prompt: "You are a Socratic tutor for introductory physics.
Never give direct answers. Instead, guide the student with questions
that help them discover the answer themselves."

Language Production Studies

Collect naturalistic language samples:

System prompt: "You are a friendly conversation partner. Discuss
everyday topics. Ask follow-up questions to keep the conversation
going. Match the participant's language complexity level."

Human-AI Interaction Studies

Research how people interact with AI systems:

System prompt: "You are a customer service agent for a fictional
airline. Help the participant rebook a cancelled flight. You have
limited options available."

Data Collected

Each conversation produces the following data:

Field	Description
turn_number	Sequential turn number
participant_message	What the participant typed
ai_response	The AI's response text
api_latency_ms	Time taken for the API to respond
tokens_used	Tokens consumed in this exchange
finish_reason	Why the response ended (STOP, MAX_TOKENS, SAFETY, etc.)

Events logged:

• chat_started — Chat session began
• message_sent — Participant sent a message
• response — AI response received
• chat_finished — Participant clicked Finish

Setup Requirements

This task requires a Gemini API key configured as a Firebase secret:

firebase functions:secrets:set GEMINI_API_KEY

Enter your Google AI Studio API key when prompted. The key is stored securely and only accessible by the Cloud Functions at runtime.

Participant-Facing Labels

The following labels can be customized per language in the task's meta-translations:

Label	Default (EN)	Description
Send	Send	Send button text
Finish	Finish	Finish button text
Placeholder	Type your message...	Input placeholder text
Typing	AI is typing...	Typing indicator text
Error	An error occurred...	Error message text
Budget exhausted	Conversation limit reached	Shown when token/turn budget is exceeded