[MODEL] all claude 5 models respond with hot take poetry

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Amodel) for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

### Type of Behavior Issue

Other unexpected behavior

### What You Asked Claude to Do

This is in response to any request of claude

### What Claude Actually Did

When responding, claude uses what I call hot take poetry.

Hot take poetry has the flavor of an online hype article and the combinations of words it uses is poem like in that it can be difficult to parse and understand what it's even trying to say.

honestly, it feels like the watermarking might be contributing to it using these odd combinations of words. it's really unmistakable when you encounter something that claude wrote.

### Expected Behavior

I expect claude's responses to be easy to digest, be succinct and delivered in plain language. Claude is a productivity tool, and it's great at autonomous work, but when a human in the loop is required, those productivity gains go out the window.

### Files Affected

```shell

```

### Permission Mode

Accept Edits was ON (auto-accepting changes)

### Can You Reproduce This?

Yes, every time with the same prompt

### Steps to Reproduce

just make any basic request of claude.

### Claude Model

Opus

### Relevant Conversation

```markdown

```

### Impact

Critical - Data loss or corrupted project

### Claude Code Version

all versions

### Platform

Anthropic API

### Additional Context

_No response_

No affected files, tests, or conversation are provided. Start by reproducing a basic request against the Anthropic API, then determine which model behavior or product component controls the response style; done means responses are consistently plain, succinct, and easy to digest without affecting normal coding behavior.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.