SocialCoach
Rehearse your conversationPractice
arXivFrom the research paper

Your personal AI social skills coach

Say the thing you’ve been not saying.

A last-minute task. A toast you want to decline. A friend who still owes you. Practise with an AI character who has a position of their own, then revisit your words and try another response.

Choose a conversation Enter a 3D scene
Have a situation in mind? Rehearse it
  • No account
  • History stays on your device
  • Chinese / English
Now with 3D scenes and video lessons
A SocialCoach 3D work dinner with three characters waiting for your responseIn the app · Work dinner (cropped) Everyone has a cup up—we’re waiting for you.

SocialCoach

Here for the practice. Here for the debrief.

Start with something on your mind

58 scenarios

Workplace

Declining Extra Hours

Ten minutes before you leave, your manager asks you to stay late — again.

Practise this conversation

Workplace

Asking for a Raise

Your project just shipped; the mid-year comp window closes this week.

Practise this conversation

Friendship

The Friend Who Hasn't Paid You Back

Three months, not a word — and they just posted vacation photos.

Practise this conversation

01Practice & reflection

After your opening line,
they’ll have another question.

Characters have positions of their own. Ask, decline, reconsider or keep talking. When you finish, revisit your own words with the coach and see what happened.

The current SocialCoach scenario collection with search and context filters
Choose an existing scene, or rehearse your own situation.
  1. 01

    Pick a scene, or bring your own

    58 bilingual scenarios across seven everyday contexts, plus custom rehearsals for the conversation ahead of you.

  2. 02

    Speak, then handle the reply

    Respond in text or use voice input. Add timed replies for more pressure, or keep the conversation going when you need to.

  3. 03

    Revisit your words. Try again.

    Read specific feedback, try another phrasing and ask the coach follow-up questions. Your practice history helps shape what comes next.

An excerpt from a demo practice

Oh… okay, let me take a look. I do have something tonight though, so I might start a bit late, is that all right?

Demo excerpt. Feedback is generated from each practice conversation.

You meant “not tonight.”
You said “a bit late.”

This debrief revisits “let me take a look” and “is that all right?” The boundary was still unclear, leaving room for more pressure. A specific line gives you something concrete to work on.

See the original debrief screenshot

The conversation with your coach continues.

Why was that line unclear? What if they keep asking? Ask follow-up questions about this practice, and look for recurring patterns across your past conversations.

023D scenes

Step into the room.
Practise with everyone listening.

Work, family and school dinners, an elevator lobby and an office. Respond to different people, choose whom to address and handle interruptions.

  • Text or voice input
  • Change perspective, move and act
  • Debrief with quotes from your conversation
Enter a 3D scene
A cropped view of three characters and the conversation in SocialCoach’s 3D elevator lobby
At the elevator: you have just left HR, with your boss and colleagues nearby.

03Watch & learn

Watch a scene.
Then take your turn.

See how different responses change the same conversation. Videos include Chinese and English captions, a breakdown of key choices and a link to the matching 3D practice.

Original fictional demonstrations · Chinese audio / Chinese and English captions

Explore 42 strategies and 30 cases

Conversation practice

Start with the conversation ahead of you

Choose a fictional practice scenario shaped around a real-life problem. Speak with a character who has a position of their own, then review what you actually said.

Workplace

Asking for a Raise: Practice the Conversation

Your project just shipped; the mid-year comp window closes this week.

View practice

Workplace

Declining Overtime: Practice Saying No

Ten minutes before you leave, your manager asks you to stay late — again.

View practice

Workplace

Called Out in a Meeting: Practice Handling Conflict

A teammate calls you out in front of the whole team for not giving feedback.

View practice

Family

Talking About Chores With a Partner: Practice the Conversation

After the hundredth time cleaning the kitchen alone, you're done keeping quiet.

View practice

Friendship

When a Friend Hasn't Repaid You: Practice Asking

Three months, not a word — and they just posted vacation photos.

View practice

04Corpus and sources

Evidence before judgment, sources before advice

58
bilingual scenarios
42
strategies
30
cases
7
context types
34
social skills

Every strategy and case carries a source.

The skill map draws on CASEL’s social and emotional learning framework.

Sources include: Difficult Conversations, Nonviolent Communication, Getting to Yes, Crucial Conversations, Never Split the Difference, Radical Candor, The Seven Principles for Making Marriage Work, Thanks for the Feedback.

05Data and deployment

Your data stays yours

History lives on your device

Practice history stays in this browser and can be exported, restored or reset. Generating content sends relevant conversation and practice context to the selected model service.

Bring your own model

Anthropic, OpenAI, or any OpenAI-compatible endpoint. Use your own API key or point it at a local model; requests go straight from your browser.

Self-hostable

The application is open source under Apache 2.0. Deploy with Docker Compose, Vercel and other platforms; the product corpus and skill map ship with the code.

06Questions

Frequently asked questions

Do I need an account or a paid plan?

No account is required. Browse scenarios and learning material directly. AI practice uses a shared model quota; if it is exhausted or unavailable, connect your own model in settings. Your provider charges for your own model usage.

How is this different from asking ChatGPT to play my manager?

A general chat model can play a role once. SocialCoach fixes the training constraints: characters have their own objectives, sometimes with an unspoken concern and won't leak the answer; text practice uses extendable segments and records unmet objectives; afterwards every point quotes the transcript first, then gives attribution, a source and the next practice.

Why won't the character give in?

Characters respond according to their goals and the conversation. Politeness does not guarantee agreement. You can negotiate, repair a misunderstanding or hold a boundary; the debrief considers communication and the outcome separately.

Where do the proficiency numbers come from?

The coach model estimates it from each conversation, and the UI says so. Everything it rests on is visible: the quoted line, the dimension, the source. Over several rounds the movement on the radar tells you more than any single number.

Where do my conversations go?

Practice history is stored in this browser and can be exported, restored or reset. Relevant context is sent to the selected model service when generating content. Anonymous usage statistics can be disabled in settings; feedback you submit is sent to the team.

Will it just write the perfect line for me?

Hints name the next move; the words are still yours. The debrief offers a better way to say it, but its job is to get you more reps under pressure.

07Research

The paper behind the product

Preprint · cs.HC · 2026

arXiv · 2606.04155

SocialCoach: Personalized Social Skill Learning with Agentic Tutoring and Practice

Tianfu Wang1, Max Xiong2, Yuxuan Lei3, Jianxun Lian4, Hongyuan Zhu3, Zhengyu Hu1, Linxiao Gong1, Dapeng Hu3, Xiaofang Li3, Peiting Tsai3, Nicholas Jing Yuan3, Qi Zhang3

1HKUST (Guangzhou) · 2Duke University · 3Microsoft · 4Microsoft Research Asia

arXiv PDF GitHub Supplementary materials BibTeX

The paper frames “what to practise next” as cold-start, retrieval-constrained sequential decision making: given a learner profile, a simulated proficiency state and the practice history, a policy writes a structured prescription that corpus retrieval then realises. The corpus itself is a traceable theory-to-practice framework that both the scheduler and the reflective tutor read from. The distinction every debrief in the app rests on, not knowing the move versus failing to execute it, comes from here too.

What the paper covers that the app can't show

  • A scheduling policy trained with trajectory-level GRPO on rubric-judge pairwise preferences
  • A synthetic cold-start evaluation against retrieval baselines and matched ablations
  • A 50-participant study with at least 10 practices each, plus a randomised pilot with n = 10
  • A 43,170-entry structured knowledge corpus built by an automated pipeline
Paper and productThe paper studies the research system and an internal research platform; the product on this site is its productised version, shipping a smaller, source-checked corpus.
Cite
@article{wang2026socialcoach,
  title   = {SocialCoach: Personalized Social Skill Learning with Agentic Tutoring and Practice},
  author  = {Wang, Tianfu and Xiong, Max and Lei, Yuxuan and Lian, Jianxun and Zhu, Hongyuan
             and Hu, Zhengyu and Gong, Linxiao and Hu, Dapeng and Li, Xiaofang and Tsai, Peiting
             and Yuan, Nicholas Jing and Zhang, Qi},
  journal = {arXiv preprint arXiv:2606.04155},
  year    = {2026}
}

Give that difficult conversation
a first try here.

Ask for a raise, bring up an unpaid loan or finally talk through something you have put off. Start with the conversation that matters to you.

Choose a conversation Have a situation in mind? Rehearse it
The SocialCoach cat coach