Master DeepSeek-V4 roleplay with custom thinking control

If you’ve been using DeepSeek-V4 for creative writing or character-based interactions, you’ve likely noticed the <think> tags. Most users treat these as a black box, but you can actually steer the model’s internal monologue to drastically improve your output quality. By injecting specific control instructions into your first prompt, you can force the model to either "live" inside the character or maintain a cold, analytical distance.

Most guides get this wrong by suggesting you put these instructions in the system prompt. In my experience, that’s a mistake. The model responds most reliably when you append these markers to the end of your very first user message. Because the model maintains the full conversation history, this initial injection acts as a persistent "mode" for the entire session.

Choosing your thinking mode

You have two primary ways to influence the model’s cognitive process. The first is Character Immersion, which forces the model to generate internal monologues within the <think> block. This is perfect for nuanced, emotional, or highly subjective roleplay. The second is Pure Analysis, which strips away the "inner voice" and forces the model to act like a director planning a scene.

Here is how these modes differ in practice:

1. Character Immersion Mode: Use this when you want the model to "feel" the scene. It forces the model to use first-person internal dialogue, such as "(I hope they don't notice I'm nervous)," which leads to more authentic, character-driven responses.
2. Pure Analysis Mode: Use this when you need structural consistency. It forbids internal monologue and forces the model to focus purely on plot progression and logical planning, resulting in a more calculated, "director-like" output.

To implement these, simply append the relevant instruction block to your first message, leaving a blank line between your prompt and the instruction. If you aren't seeing the desired behavior, don't worry—it’s probabilistic. Just start a new chat and try again.

Why this works for power users

The magic happens because you are essentially priming the model’s latent space before it generates the first token. By explicitly defining the rules for the <think> tag, you are constraining the "thought" process to follow a specific pattern. This is a classic prompt engineering technique that separates casual users from those who actually understand how to manipulate LLM reasoning.

Here is the specific instruction block for Character Immersion: 【角色沉浸要求】在你的思考过程（<think>标签内）中，请遵守以下规则：1. 请以角色第一人称进行内心独白，用括号包裹内心活动。2. 用第一人称描写角色的内心感受。3. 思考内容应沉浸在角色中，通过内心独白分析剧情和规划回复。

If you prefer the Pure Analysis approach, swap that for: 【思维模式要求】在你的思考过程（<think>标签内）中，请遵守以下规则：1. 禁止使用圆括号包裹内心独白。2. 禁止以角色第一人称描写内心活动。3. 思考内容应聚焦于剧情走向分析和回复内容规划。

This next part matters more than it looks: if you want to switch modes mid-conversation, you can't just edit the prompt. You must start a fresh chat. The model’s "thinking" style is locked in by the initial context window, so treat your first message as the foundational architecture for the entire roleplay session.

Mastering these DeepSeek-V4 roleplay instructions allows you to move beyond generic responses and start crafting truly immersive, character-driven narratives. Try this today and share what you find in the comments.