AI Twin Chat¶

The Chat screen connects you to MyTwin — an AI that speaks as an insightful, grounded version of yourself, not as a generic assistant. It is powered by the Claude API with streaming responses.

How It Works¶

sequenceDiagram
    participant User
    participant ChatViewModel
    participant ChatRepository
    participant ClaudeContextBuilder
    participant SimulationRepository
    participant ClaudeApiService
    participant Claude API

    User->>ChatViewModel: sendMessage()
    ChatViewModel->>ChatRepository: streamResponse(messages)
    ChatRepository->>SimulationRepository: getBundle()
    ChatRepository->>ClaudeContextBuilder: build(profile, bundle, history)
    ClaudeContextBuilder-->>ChatRepository: ClaudeContextSnapshot (systemPrompt)
    ChatRepository->>ClaudeApiService: streamMessage(systemPrompt, messages)
    ClaudeApiService->>Claude API: POST /v1/messages (SSE)
    Claude API-->>ClaudeApiService: content_block_delta tokens
    ClaudeApiService-->>ChatViewModel: Flow<String> (token by token)
    ChatViewModel-->>User: live bubble update

System Prompt¶

ClaudeContextBuilder assembles the system prompt from three sources:

1. Persona instructions¶

You are MyTwin, the user's more insightful digital twin.
Speak like an intelligent future version of the user, not like a generic AI assistant.
Sound grounded, calm, practical, and a bit personal without becoming theatrical.
Keep answers concise by default. Usually answer in 2 to 5 short paragraphs or compact bullet points.

2. Anonymized user context¶

Age, biological sex, height, weight
Smoking, alcohol, diet quality
Average sleep and steps baselines
Perceived stress level

No name is included in the context — "digital twin" feels less clinical when it speaks from a shared perspective rather than about "you".

3. Simulation baseline + what-if scenarios¶

The full SimulationReport (scores, confidence, narrative) and all five precomputed scenario comparisons are injected as structured text. This lets Claude answer questions like "what would happen if I slept more?" with grounded, data-backed answers rather than generic advice.

4. Recent wearable history¶

7-day day-by-day summaries for sleep, steps, heart rate, and stress — formatted as YYYY-MM-DD:value pairs.

Streaming¶

ClaudeApiService opens a raw HttpURLConnection with:

Accept: text/event-stream

It parses SSE events line-by-line:

event: content_block_delta
data: {"delta": {"type": "text_delta", "text": "Your "}}

event: content_block_delta
data: {"delta": {"type": "text_delta", "text": "recovery "}}

Each text token is emitted on a Flow<String> on Dispatchers.IO. ChatViewModel collects this flow and appends tokens to the current ChatMessage.text in StateFlow, so the bubble updates in real time as Claude types.

API Configuration¶

Setting	Value
Endpoint	`https://api.anthropic.com/v1/messages`
Model	Configured via `local.properties` (`claudeModel`)
Max tokens	700
Temperature	0.6
Stream	`true`

The API key (claudeApiKey) is injected at build time via BuildConfig from local.properties. It is never committed to source control.

UI¶

Message list — a LazyColumn that auto-scrolls to the latest message
Streaming indicator — while Claude is responding, the bubble shows a "typing…" label below the growing text
Input field — multi-line OutlinedTextField, disabled while a response is streaming
Error card — if the API call fails, a dismissable error card appears above the input

Safety Note¶

The system prompt explicitly frames this as a coaching-style wellbeing projection, not a diagnosis:

If a question sounds urgent or medical, say professional care is appropriate.

Claude will not attempt to diagnose conditions or prescribe treatments.