Social Confidence Coach: A Private AI Practice Partner for My Roommate

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

I built the Social Confidence Coach for my roommate, who struggles with social anxiety, meeting new people, and initiating casual conversations.

Approaching peers or making small talk can feel intimidating and stressful. I wanted to give him a zero-stakes, completely private simulation where he can select different difficulty levels (from low-pressure barista chats to campus conversation starters), practice what to say, and receive real-time constructive feedback from an AI coach on whether his responses are engaging, natural, and polite.

file:///private/var/folders/zz/ms_5j6rx5q195c7yt3x7k_y40000gn/T/TemporaryItems/NSIRD_screencaptureui_jfvbK3/Screenshot%202026-10-04%20at%206.57.28%E2%80%AFPM.png

python
import ollama

def main():
    print("=== Social Confidence Daily Challenge & AI Practice ===")
    print("Choose your practice scenario:")
    print("1. Low Pressure - Casual greeting with a barista/cashier")
    print("2. Medium Pressure - Small talk with a student at a campus cafe")
    print("3. Active Engagement - Chatting with someone about shared interests")

    choice = input("\nEnter 1, 2, or 3: ").strip()

    scenarios = {
        "1": "You are a friendly barista at a local cafe during a quiet afternoon.",
        "2": "You are a fellow student sitting at a campus cafe studying for an exam.",
        "3": "You are an approachable student sitting in the library reading an interesting book."
    }

    selected_scenario = scenarios.get(choice, scenarios["2"])

    system_instruction = (
        f"Roleplay scenario: {selected_scenario} "
        "Keep your in-character answers brief (1 to 2 sentences). "
        "On a brand new line, always write '[Coach Advice:]' and give 1 brief sentence "
        "evaluating whether the user's message was natural, open-ended, or polite."
    )

    messages = [{"role": "system", "content": system_instruction}]

    print("\nScenario started! Type 'exit' to quit.\n")

    while True:
        user_input = input("You: ")

        if user_input.strip().lower() == "exit":
            print("\nSession complete. Great practice!")
            break

        messages.append({"role": "user", "content": user_input})

        response = ollama.chat(
            model="llama3.2:1b",
            messages=messages
        )

        reply = response["message"]["content"]
        print(f"\n{reply}\n")

        messages.append({"role": "assistant", "content": reply})

if __name__ == "__main__":
    main()

## How I Built It
The application is built in Python and runs locally using the official ollama Python library and Meta’s open-weight Llama 3.2 1B model.

Scenario Selection: The user chooses a comfort level, which dynamically constructs a tailored system prompt for the roleplay.

Context Memory: The script maintains a conversation history array (messages), ensuring the open-source model retains context across multi-turn interactions.

Dual Role Output: The model is instructed to output both an in-character peer dialogue and an analytical [Coach Advice:] snippet after each turn.

## Why Does Open Innovation Matter?
Open innovation and open-weight models are vital for a tool like this:

Total Privacy for Sensitive Topics: Practicing social interactions, dating openers, and overcoming social anxiety feels deeply personal. Closed cloud APIs store chat logs on remote servers. Running Llama 3.2 locally on a laptop via Ollama ensures that sensitive practice conversations never leave the machine.

Offline & Zero Cost: Social self-improvement tools should not require a recurring credit card charge or active internet connection. Open-source AI makes this available anywhere, completely free of charge.

## My Agent Session
I developed this project step-by-step with the help of an interactive AI coding assistant. 

During the build process, the assistant helped me:
1. Set up the local environment and troubleshoot the `ollama` library installation in VS Code.
2. Structure the conversation memory loop in Python so the model remembers chat context across multiple turns.
3. Engineer the dual-response system prompt so the local open-source model reliably outputs both in-character dialogue and actionable coaching advice in every turn.

##Feedback from My Friend
I had my roommate sit down and test it out by picking Level 2. After trying a couple of openers, his reaction was:

"It actually feels way less intimidating to test lines here than feeling awkward in person, and the quick tips tell me if I sounded too blunt."

## Prize Categories
Primary Track: Build for a Friend

<!-- Thanks for participating! -->