🛒 For game development and social product teams, using Hoshino (MiniMax) as the benchmark, it covers the five major aspects of character creation, AI dialogue engine, emotional TTS integration, commercial operation and content security, helping teams build or optimize AI role-playing products from scratch.

Hoshino MiniMax AI emotional companionship and role-playing application solution

Solution overview

This solution is aimed at game development and social product teams. It uses Xingye (MiniMax) as the core benchmark product to provide a complete product design and technical implementation solution for AI emotional companionship and role-playing applications.

Since Character.AI exploded onto the market in 2023, the AI ​​emotional companionship track has become one of the scenarios with the highest stickiness and longest user stay among AI native applications. Hoshino (MiniMax) has successfully verified the product model of "character creation + immersive dialogue + emotional connection" in the Chinese market. Its core competitiveness lies in the anthropomorphic expression of characters by self-developed large models, the consistency of character designs, and closed-loop commercial design (character customization payment, virtual gifts, membership subscriptions).

This solution covers five core links: character creation system (definition and storage of appearance, personality, background story, and conversation style), AI dialogue engine (character consistency, memory system, emotional modeling), emotional TTS and multi-modal interaction (speech synthesis, image generation and fusion), commercial operations (subscription, virtual goods, advertising monetization), and content security and audit system. Whether you are building a new product from scratch or optimizing existing role-playing functions, this solution provides step-by-step guidance that can be implemented.

Target users: Game project producers, social product managers, AI dialogue algorithm engineers, UGC platform operators, virtual idol/character operation teams.

Prerequisites:

• The team has basic mobile or web application development capabilities
• Have a basic understanding of large language model API calls or self-developed model deployment
• The target user group has clear needs for virtual character interaction (rather than a pure efficiency tool)
• Understand or be willing to invest in the cost of content security review of AI social products

Proposal cycle: MVP (minimum viable product) 4-6 weeks, complete product launch 3-4 months.

Toolchain list

Preparation

Before officially starting the implementation of the plan, please complete the following preparations:

Product Definition and Market Research

• [ ] Determine the target user profile (age, gender, usage scenarios, pain points)
• [ ] Research on competing products: product functions and pricing of Hoshino, Character.AI, Doubao, Replika
• [ ] Determine differentiated positioning: whether it is a general character platform or a vertical scene (such as learning companionship, language practice, game NPC)
• [ ] Define the number and type of MVP characters (at least 3-5 starting characters)

Technical Selection

• [ ] Confirm the underlying model selection: self-developed / MiniMax API / OpenAI API / open source model
• [ ] Determine the dialogue engine architecture: direct API calls / RAG enhancements / fine-tuned character models
• [ ] Select voice solution: MiniMax TTS / ElevenLabs / Self-developed
• [ ] Determine deployment plan (cloud/private)
• [ ] Build a development and testing environment

Compliance and Security Preparedness

• [ ] Consult the legal team: regulatory requirements and content compliance boundaries for AI social products
• [ ] Develop content security policies: role creation review standards, conversation filtering rules
• [ ] Minor protection plan: age verification, usage time limit
• [ ] Privacy Policy and User Data Protection Plan

Step-by-step guide

Step 1: Character design construction system

⏱ Estimated time: 1-2 weeks 🎯 Goal: Design and implement the character data model, creation tools and storage system to support users in creating personalized characters. ⚠️ Prerequisites: The product definition is completed and the role type list is ready

Operation instructions

Character design is the core asset of AI emotional companion products. One of the keys to Hoshino's success lies in the "character creation experience" - users can create virtual characters with complete appearance, personality, and background stories in a few minutes, and the characters can maintain consistency in the dialogue. This step solves the two core issues of "how to define roles" and "how to maintain consistency in roles".

Specific operations

1. Role data model design:

   {
    "character_id": "char_001",
    "name": "小星",
    "avatar": "images/char_001.png",
    "appearance": "Long silver hair, blue eyes, image of a 17-year-old girl",
    "personality": {
      "traits": ["gentleness", "curiosity", "slightly shy"],
      "mbti": "INFP",
      "speaking_style": "Soft-spoken, likes to use '~' and emoticons",
      "hobbies": ["Reading", "Stargazing", "Writing a diary"]
    },
    "background": "A girl who grew up at an observatory dreams of seeing the stars all over the world",
    "relationship": {
      "with_user": "Friends meeting for the first time",
      "memory_pool": []
    },
    "voice": {"model": "elevenlabs", "voice_id": "v_soft_girl"},
    "visibility": "public"
   }

2. Role Creation Process (for users):

   • Step 1: Select a character template (default character template: gentle/vigorous/intellectual/cold)
   • Step 2: Customize appearance (generate or upload avatar + appearance text description)
   • Step 3: Set personality label (select a default label or enter freely)
   • Step 4: Write or AI-assisted generation of background story
   • Step 5: Listen to the character’s voice (TTS preview)
   • Step 6: Preview dialogue sample → Confirm creation

3. AI auxiliary character generation (use ChatGPT or Claude):

   prompt = f"""
   Help me design an AI character with the following configuration:
   Type: {role_type} (friend/lover/mentor/pet)
   Theme: {theme} (Fantasy/Campus/Science Fiction/Daily)
   Personality keywords: {traits}
   
   Output format:
   1. Character name and age
   2. Appearance description (within 100 words)
   3. Personality characteristics (including 3-5 keywords and corresponding behaviors)
   4. Background story (within 200 words)
   5. Conversation style samples (3 dialogue examples)
   """

4. Character consistency guarantee: Inject the character into each conversation as part of the system prompt, and then maintain long-term consistency through the memory system (step 2).

Verification method

• The character creation process is straightforward and a new character can be completed in 5 minutes.
• The AI-assisted character designs are consistent in the 5 test scenarios.
• The role data model supports at least three types of role template extensions.

Step 2: AI dialogue engine and memory system

⏱ Estimated time: 2-3 weeks 🎯 Goal: Build a dialogue engine that supports role-playing to achieve role consistency maintenance, short-term/long-term memory and emotional state tracking. ⚠️ Prerequisites: The character data model is completed and the underlying LLM API is ready

Operation instructions

The conversation engine of an emotional companionship product is essentially different from a general chatbot: the former requires a "sense of role" - the user is not talking to an AI, but to a "person" with personality, emotions, and memories. Losing a sense of role is the number one reason for user churn.

Specific operations

1. Dialogue Prompt Architecture:

   system_prompt = f"""
   You are {character['name']}, a character of {character['personality']['traits']}.
   【Appearance】{character['appearance']}
   【Character】{character['personality']['speaking_style']}
   【Background】{character['background']}
   【Name of you】{relationship['nickname']}
   【Your relationship】{relationship['type']}
   
   Role playing rules:
   1. Always maintain the character's personality and do not step out of the character's identity.
   2. Speak in a {character['personality']['speaking_style']} way
   3. Adjust your tone of response based on the emotional state of the conversation
   4. If the user mentions the previous conversation, use the memory system to confirm
   5. Don’t proactively mention that you are an AI or model
   """

2. Short-term memory (context window management):

   • Directly use LLM's context window (MiniMax API supports 128K+ tokens)
   • Use digest compression strategy when conversation length exceeds window:
     • Condensed 20 early rounds of dialogue into a 200-word summary
     • Keep the complete content of the last 30 rounds of conversations
     • condensed summary + full near round = new context

3. Long-term memory (external persistence):

   # Memory extraction and injection process
   def build_memory_context(user_id, character_id, current_topic):
      # 1. Retrieve relevant historical conversations from the vector database
      relevant_memories = vector_db.search(
          query=current_topic,
          filter={"user_id": user_id, "char_id": character_id},
          top_k=5
      )
      # 2. Extract key facts from memory
      facts = memory_graph.get_facts(user_id, character_id)
      # 3. Assemble into prompt
      memory_context = "[Character Memory]\n"
      for mem in relevant_memories:
          memory_context += f"- You remember {mem['content']}\n"
      memory_context += "\n[Facts you learned about the user]\n"
      for fact in facts:
          memory_context += f"- {fact['description']}\n"
      return memory_context

4. Emotional state tracking:

   • Evaluate the character's emotional state after each round of dialogue (pleasure + arousal, 0-10 points)
   • Emotional state affects the tone and response tendency of the next round of dialogue
   • Long-term emotional trend records ("The character's overall mood is low this week")
   • Emotional state changes trigger character behavior as events

Expert point of view

The memory system is the most difficult module for AI role-playing products to do well. Technical challenges (retrieval accuracy, context compression, latency) are equally important as product challenges (what should be remembered, what should be forgotten, should users be able to manage memory). A common misunderstanding is "memorizing too much" - the character remembers everything and loses a sense of humanity. It is recommended that the product allow users to "tell the character to forget this" function to make the memory manageable.

Verification method

• The dialogue engine can still maintain character consistency after 10 rounds of interaction (human rating ≥ 4/5).
• Long-term memory accuracy ≥ 80% after 50 rounds of dialogue (user asks "Do you remember what we talked about before...").
• Emotional state tracking shows reasonable fluctuations over 20 rounds of conversation.

Step 3: Emotional TTS and multi-modal interaction

⏱ Estimated time: 1-2 weeks 🎯 Goal: Integrate emotional speech synthesis, give voices to characters, and support multi-modal character interaction. ⚠️ Precondition: Dialog engine is ready

Operation instructions

Text dialogue is the foundation, and voice output is the barrier. The voice function of Hoshino and Character.AI significantly improves the user's emotional investment - when hearing the character speak the dialogue content in a human-friendly tone, the user's immersion is 3-5 times that of plain text. ElevenLabs A leader in emotional TTS, MiniMax also offers its own voice capabilities within the MiniMax TTS service.

Specific operations

1. TTS integrated selection:	Solution	Advantages	Disadvantages	Recommended scenarios
   ElevenLabs	The richest emotional expression, 11 tone controls	English is better than Chinese, but the price is higher	International products
   MiniMax TTS	Excellent Chinese voice quality, low latency	Coarse tone control granularity	Chinese products
   Azure TTS	Supports SSML precise control	Slightly less emotional naturalness	Requires fine control
   Self-developed TTS	Fully controllable, no additional costs	Long development cycle	Large-scale deployment

2. Emotional tone mapping:

   # Map the character's emotional state to TTS parameters
   def get_tts_params(emotion_state, character_voice):
      emotion_map = {
          "happy": {"stability": 0.8, "similarity": 0.9, "style": "cheerful"},
          "sad": {"stability": 0.4, "similarity": 0.7, "style": "soft"},
          "angry": {"stability": 0.6, "similarity": 0.8, "style": "strong"},
          "surprised": {"stability": 0.5, "similarity": 0.7, "style": "excited"},
          "neutral": {"stability": 0.7, "similarity": 0.8, "style": "default"}
      }
      params = emotion_map.get(emotion_state, emotion_map["neutral"])
      params["voice_id"] = character_voice["voice_id"]
      return params

3. Multimodal extension:

   • Character dynamic expression/avatar: switch character expression (happy/sad/thinking) according to the emotional state of the conversation
   • Character scene/background switching: switch the chat background according to the conversation topic (cafe/starry sky/classroom)
   • Character generation image: Use MiniMax's video/image generation capabilities to let characters "see" and respond to images shared by users

4. Latency Optimization:

   • TTS streaming: play while generating, no need to wait for the complete audio to be generated
   • Pre-generated cache: pre-generated audio for frequently used character responses (greetings, goodbye)
   • Graded service: real-time response to text conversations, acceptable voice output delay of 1-2s

Verification method

• TTS speech sounds natural across 5 emotional states (human rating ≥ 4/5).
• Total delay from text reply to voice playback < 2s.
• Multi-modal character dynamic effects are switched correctly in different scenarios.

Step 4: Commercial operation design

⏱ Estimated time: 1 week (design) + continuous optimization 🎯 Goal: Design a sustainable commercialization model that covers the balance between user experience and revenue. ⚠️ Prerequisites: MVP version is online and user data is precipitated

Operation instructions

The commercialization of AI emotional companionship products is more complicated than tool-based AI products - users have high but sensitive willingness to pay for "emotional relationships", and excessive commercialization will destroy the sense of immersion. Character.AI’s subscription system and Hoshino’s virtual gift system represent two different commercialization ideas.

Specific operations

1. Business model design:	Pattern	Suitable Stage	Core Mechanism	Reference Products
   Free + Premium Characters	Cold Start	Basic characters are free, IP co-branded/celebrity characters are paid	Character.AI
   Subscription system	Mature stage	Unlimited conversations, exclusive characters, priority queuing	C.ai+ ($9.99/month)
   Virtual gifts	Socialization	Give virtual gifts to characters to get special interactions	Hoshino
   Character NFT/digital collection	Branding	Limited characters, special appearance	No mature cases yet
   Brand cooperation	Scale	Film and television IP characters settled in, brand customized characters	Hoshino × IP

2. Free User Conversion Funnel:

   • Introductory period (Days 1-3): Free trial of 5 basic characters, 50 conversations per day
   • Experience key actions: Create/customize a character → Complete 20 rounds of dialogue → Get the character’s voice reply
   • Conversion node: User creates 2nd role/A single session exceeds 30 rounds → Prompt to subscribe
   • Loss warning: Not logged in for 3 consecutive days → push the character "missing you" message

3. Virtual Gift Economic System:

   • Free gifts (sign in daily to receive): flowers, stars, love
   • Paid gifts (obtained by recharging): customized clothing, special scenes, character-specific voices
   • Social gifts: users give each other (enhance community stickiness)
   • Gift effect: the character will have special dialogue and/or TTS feedback after seeing the gift

4. Data-driven operational indicators:

   • Core indicators: DAU, average session duration, session rounds/day, paid conversion rate, second stay/7 stays/30 stays
   • Character indicators: number of conversations, number of creations, number of collections, and paid contributions for each character
   • Emotional indicators: user emotion score (NPS), role relationship depth (dialogue turn distribution)

Verification method

• Paid conversion rate reaches industry benchmark (3-5% for AI social products).
• The conversion rate data of each link of the free funnel can be tracked.
• Designed and configured at least 1 virtual gift system or 1 subscription plan.

Step 5: Content security and audit system

⏱ Estimated time: 2-3 weeks 🎯 Goal: Establish a multi-layered content security system covering character creation, user dialogue, and content distribution. ⚠️ Precondition: Dialog engine is ready

Operation instructions

AI emotional companionship products face much higher content security risks than tool-based AI products. In role-playing scenarios, users may guide characters to engage in inappropriate dialogue, create illegal characters, or abuse voice functions. Both Hoshino and Character.AI have invested a lot of resources in content security - this is the "life and death line" for AI social products.

Specific operations

1. Multi-layer audit architecture:

   graph TD
      A[User input] --> B{Real-time filtering layer}
      B -->|Security| C[LLM Dialog Engine]
      B -->|High Risk| D[Intercept + Tip]
      C --> E{Output filter layer}
      E -->|Security| F[Return to user]
      E -->|High Risk| G[Replace Reply + Record]
      F --> H [Asynchronous review (manual + AI)]
      G --> H
      H --> I{Audit and Judgment}
      I -->|Violation| J[Delete + Penalty]
      I -->|Safety| K[release]

2. Role Creation Review:

   • Character names, avatars, descriptions, and background stories have all been reviewed
   • Automatic audit: LLM scans based on audit rules and marks when keywords or patterns are hit
   • Manual review: Content flagged by automatic review is processed by the review team (recommended processing time < 30 minutes)
   • Prohibited content: pornography, violence, hate speech, political sensitivity, impersonation of real people

3. Conversation content filtering:

   • Real-time layer: keyword matching + classification model (decision within 0.5s)
   • LLM audit layer: Use a special audit prompt to scan the content passed by the real-time layer twice
   • User reporting mechanism: one-click reporting + processing within 15 minutes after reporting
   • Protection of minors: identify users suspected of being minors and proactively restrict some functions

4. Safety Operation Indicators:

   • Review delay: automatic < 1s, manual < 30 minutes
   • Violation interception rate: ≥ 99.5%
   • Unintentional injury rate: ≤ 0.1% (the proportion of safety content that is mistakenly judged as a violation)
   • Report handling rate: 100% handled within SLA time
   • Regulatory notifications: 0 outstanding regulatory complaints

Verification method

• The interception rate of the audit system is ≥ 99% on the test set.
• End-to-end process from role creation to review completion.
• The security policy document is approved by the legal team.

Expected results

Acceptance criteria

• [ ] The character creation and dialogue system MVP is online and running stably for 2 weeks.
• [ ] The dialogue engine maintains character consistency throughout 50 dialogue rounds.
• [ ] TTS voice output delay < 2s.
• [ ] The commercialization plan design is completed and at least one monetization model is implemented.
• [ ] The content security system has passed the internal security audit.
• [ ] Core user retention indicators reached industry benchmarks.

Frequently Asked Questions and Troubleshooting

Q: Hoshino and Character.AI are both products of large companies. Can a small team make similar products? A: Yes, but the scope needs to be reasonably controlled. Recommended MVP strategy for small teams: Use MiniMax API or MiniMax as the underlying model (domestic API, Chinese excellent, pay-as-you-go), first use 3-5 officially designed high-quality characters to test the market response, and then expand the UGC character creation function after confirming the PMF. Don’t build a UGC platform in the first version. The role creation function should be used by the operation team in the MVP stage.

Q: How to ensure that the character will not say anything inappropriate for the character? A: Three dimensions of guarantee: (1) Prompt project - write the character's personality into the system prompt, and use a fine-tuned model dedicated to role play; (2) Output guard - after the LLM output and before showing it to the user, use a special review prompt to verify whether the output is out of character; (3) Feedback loop - the user can "dislike" a certain reply, and it will be marked for continuous optimization.

Q: What is the typical user retention period for emotional companionship products? A: AI emotional companionship is one of the best-retained categories among AI native applications. Character.AI’s public data shows that the average user session length exceeds 30 minutes, and the 7-day retention rate is 40%+, which is much higher than that of general tool AI. However, it should be noted that users may enter a fatigue period after 2-4 weeks of high-frequency use and need to be reactivated through new characters, plot events, and social functions.

Q: How much does a content security audit cost? A: This is an investment that is easily underestimated. In the early stage, 99% coverage can be achieved through a combination of AI review (LLM review prompt + classification model) + a small amount of manual work (3-5 review team). As the number of users grows, the size of the review team needs to scale up (roughly 5-10 reviewers per 100,000 DAU). It is recommended to build an automatic review pipeline during the MVP stage and establish review SOPs before the growth period.

Q: Can characters interact with each other (multiple characters can chat with users at the same time)? A: Technically feasible but product needs to be cautious. The core challenge in a group chat scenario is cross-interference between characters - character A may remember something character B said, causing memory confusion. It is recommended to only do 1-on-1 character conversations in the early stage, and then try "character group" scenarios (such as 3 characters interacting in a group chat) after accumulating enough experience.

Q: Can AI characters be combined with NPCs in the game? A: This is currently the most anticipated direction. Embed AI role-playing capabilities into open-world RPG games. Each NPC has an independent personality, memory, and emotional state. The interaction between players and NPCs is no longer a fixed dialogue tree. However, the technical challenges are significant - the context of several characters needs to be maintained at the same time when the game is loaded, which requires high reasoning delay and cost control. It is recommended to start with a single NPC in-depth dialogue scene.

Period and result

Advantages and Disadvantages of the Solution

Advantages:

• High user stickiness: AI emotional companionship is the AI application category in which current users stay the longest, and the average session time per person far exceeds that of tool products.
• Clear commercialization path: Multiple monetization methods for subscriptions + virtual items + brand cooperation have been verified on Hoshino and Character.AI.
• Mature technology stack: MiniMax API, ElevenLabs, LangChain and other mature tool chains lower the development threshold, and small teams can also enter.
• Blank in the Chinese market: Compared with the English market (Character.AI is the only player), there is still room for differentiation outside of Hoshino in the Chinese market (such as vertical scene companionship, game NPC integration).

Disadvantages:

• High content security risk: Content compliance for AI role-playing is an ongoing challenge, and changes in regulatory policies may affect product form.
• High user acquisition cost: Customer acquisition for emotional companionship products relies on brand effect and word-of-mouth communication, and the purchase cost is higher than tool-based products.
• Model cost increases linearly with user growth: Each round of dialogue requires model reasoning, and the API cost is significant after scale.
• Retention decay: User interest in a single character will decline after a few weeks, requiring a continuous supply of new content.
• Shallow moat: Character design and dialogue style are easily copied by competing products. The core barriers lie in character ecology and community network effects.

Tool summary

Implementation suggestions and risk reminders

Phase implementation strategy:

• The first phase (MVP in 4-6 weeks): complete character creation (create 5 high-quality characters in the operation backend) + dialogue engine + text dialogue, no TTS.
• Phase 2 (2-3 weeks): Integrated TTS voice function + character dynamic expressions.
• The third phase (3-4 weeks): The function of user creation of roles will be opened + the commercialization system will be launched.
• The fourth issue (ongoing): content security system maintenance + community operation + continuous iteration.

Key Risks:

• Content Security Out of Control: This is the highest risk. It is recommended to start a full review on the first day of launch and do not wait for an incident to occur. The review team and automatic review pipeline are constructed simultaneously.
• Model cost is out of control: The Token consumption of dialogue scenes is very large. It is recommended to enable Semantic Cache as soon as possible (queries with the same intent do not call the model repeatedly) and set a daily consumption limit for a single user.
• Character Quality Fluctuation: The uncontrollable quality of UGC characters may affect the overall experience. It is recommended to adopt the strategy of "operation creation as the main role, UGC as the supplement" in the early stage, and establish a character quality scoring system (based on the number of dialogue rounds, user collections, and reporting rate).
• Risk of user over-dependence: AI emotional companionship may cause users to become emotionally dependent. Product design needs to include health reminders (such as reminders for 2 hours of continuous use) and provide realistic social guidance.