# Design chat messages

> Build clear text conversations within Router's supported message and context limits.

## Use roles deliberately

| Role        | Use it for                                                           |
| ----------- | -------------------------------------------------------------------- |
| `system`    | Stable instructions, tone, boundaries, and response format guidance. |
| `user`      | The current request and user-provided context.                       |
| `assistant` | Earlier model replies needed to continue a conversation.             |

Router accepts between 1 and 16 messages. Message content must be text, and the combined UTF-8 message content cannot exceed the edition’s `max_prompt_bytes` (at most 196,608 bytes). Bytes only bound the request size: input that exceeds the edition’s `context_length` in tokens is rejected with `context_length_exceeded`, never truncated.

## Keep durable instructions stable

Put product-wide behavior in one concise system message. Keep changing user context in user messages. This makes a conversation easier to inspect and reduces duplicated prompt text.

```json
{
  "model": "gemma3:1b",
  "messages": [
    {
      "role": "system",
      "content": "You are a concise product support assistant. Say when information is missing."
    },
    {
      "role": "user",
      "content": "Explain the difference between online providers and free capacity."
    }
  ]
}
```

## Manage conversation history

Send only the history the next answer needs. For longer conversations, summarize older turns in your own application before the request reaches Router. Preserve facts, decisions, and unresolved questions in the summary.

## Control output length

Set `max_tokens` to the smallest useful ceiling. It defaults to 512, Router’s platform ceiling is 2,048 output tokens, and the selected model edition can be lower. Input and output share the context window, so a long prompt leaves less room for output. A response ending with `finish_reason: "length"` reached `max_tokens` or filled the context window and may be incomplete.

## Choose a temperature

`temperature` accepts values from 0 through 2. Lower values usually suit extraction, support, and repeatable formatting. Higher values permit more variation. The exact effect depends on the provider runtime and model.

> **Bytes and tokens are different**
> The request has both a UTF-8 byte limit and a model context-token limit. Non-ASCII text can use several bytes per character, so check the documented model limits before sending large inputs.

Next, learn how to [stream the answer](https://router.optimai.network/docs/streaming.md) or inspect the [chat-completions request schema](https://router.optimai.network/docs/api/chat-completions.md).