切换角色扮演的思考风格

Notes

• This document describes the special control instructions for DeepSeek-V4 roleplay, used to switch the chain of thought (CoT) style in thinking mode

• Scope of Application: Expert Mode on DeepSeek official APP / web client, as well as the API for deepseek-v4-flash and deepseek-v4-pro. Quick mode on the web version is currently not supported

• Probabilistic output: 100% triggering is currently not guaranteed, but it stably increases the probability of getting the desired format. If it doesn't work the first time, roll a few more times

Three Modes

Effect comparison (examples, not real output):

Character Immersion — "in character" like an actor:    Pure Analysis — calm planning like a director:
<think>                                                  <think>
(He greeted me... my heart is racing.)                   Scene: user greets, character is tsundere.
I'll pretend not to care.                                Reply strategy: act aloof first, body language betrays.
(Don't let him see I'm happy!)                           150 chars, action first then dialogue.
</think>                                                 </think>

Instructions (copy-paste ready)

Character Immersion mode:

【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，例如"（心想：……）"或"(内心OS：……)"
2. 用第一人称描写角色的内心感受，例如"我心想""我觉得""我暗自"等
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

Pure Analysis mode:

【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，例如"（心想：……）"或"(内心OS：……)"，所有分析内容直接陈述即可
2. 禁止以角色第一人称描写内心活动，例如"我心想""我觉得""我暗自"等，请用分析性语言替代
3. 思考内容应聚焦于剧情走向分析和回复内容规划，不要在思考中进行角色扮演式的内心戏表演

How to use on the web client

Only 1 step: paste the instruction at the end of your first message, then chat normally.

「我推开咖啡店的门，看到你正在擦吧台。」"你好，请问还有位置吗？"

【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，例如"（心想：……）"或"(内心OS：……)"
2. 用第一人称描写角色的内心感受，例如"我心想""我觉得""我暗自"等
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

Why it works: the model sees the full conversation history on every reply, so the first-round instruction stays in context and applies throughout.

Tips:

• Want to switch modes? Start a new conversation and paste the other instruction in the first message

• Don't want it? Add nothing — the model picks the most suitable thinking style

• Click「查看思考过程」to verify the mode is active

API developer reference (Python)

INNER_OS_MARKER = (
    "\n\n【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：\n"
    "1. 请以角色第一人称进行内心独白，用括号包裹内心活动，例如\"（心想：……）\"或\"(内心OS：……)\"\n"
    "2. 用第一人称描写角色的内心感受，例如\"我心想\"\"我觉得\"\"我暗自\"等\n"
    "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"
)
NO_INNER_OS_MARKER = (
    "\n\n【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：\n"
    "1. 禁止使用圆括号包裹内心独白，例如\"（心想：……）\"或\"(内心OS：……)\"，所有分析内容直接陈述即可\n"
    "2. 禁止以角色第一人称描写内心活动，例如\"我心想\"\"我觉得\"\"我暗自\"等，请用分析性语言替代\n"
    "3. 思考内容应聚焦于剧情走向分析和回复内容规划，不要在思考中进行角色扮演式的内心戏表演"
)

def build_messages(system_prompt, user_first_message, mode="default"):
    if mode == "inner_os":
        user_first_message += INNER_OS_MARKER
    elif mode == "no_inner_os":
        user_first_message += NO_INNER_OS_MARKER
    return [
        {"role": "system", "content": system_prompt},
        {"role": "user",   "content": user_first_message},
    ]

messages = build_messages("You are a tsundere high school girl...", "「I walk into the classroom」\"Morning.\"", mode="inner_os")
response = client.chat(messages)

messages.append({"role": "assistant", "content": response})
messages.append({"role": "user", "content": "「I sit down next to her」\"Are you feeling down today?\""})
response = client.chat(messages)  # Marker from round 1 is still in history, auto-applies

FAQ

Q: Can the instruction go in the system prompt? A: Recommended placement is at the end of the first user message — that's the training-time injection position and works most reliably.

Q: Does the final reply change after adding the instruction? A: The instruction only affects the thinking process, but the thinking style indirectly shapes the reply — character immersion feels more emotional, pure analysis is structurally steadier.

Other CoT modification method (lucky draw, not specially trained)

• Add to the first-round instruction: your thinking output should start verbatim with ` thinking (write the desired CoT opening here, e.g. 嗯/好的)`, output thinking only once, do not repeat ` thinking

• thinking is the fixed <think> token; the principle is changing the reasoning start character to force the model into different CoT patterns (QA, writing, reasoning, agent), but these are not trained specifically for roleplay, so results may be luck-based

Source note: this English guide was translated from the Chinese original (author Deli Chen, a DeepSeek employee) and sha… 以上为必要节选，完整内容请查看原文。