Following up on the previous article DeepSeek V4 Novel Writing Capability Review (Part 1), this article will further explore DeepSeek V4's ability to write novels.

First, we need to look at DeepSeek's ability to follow skills. Skills are a crucial aspect of FeelFish Agents, and large models need to be able to use and adhere to these skills effectively to meet our creative needs.

We started a new novel project and tried adding an opening skill:

After adding the skill, we asked the Agent to start the novel, using the DeepSeek V4 Flash thinking mode for testing:

The skill was successfully invoked, and the thought process is as follows:

Thought Process
Collapse
The user wants to create a web novel, with the project name "Opening a Restaurant in Tang Dynasty" and the pen name "Foolish Guide." This is a new project with no existing files.

I need to guide the user step by step:
1. Understand the user's creative intentions and preferences - through dialogue
2. Help build the overall outline of the novel
3. Define the main character settings
4. Establish the overall style and creation rules of the novel
5. Finally, create the content for the first chapter

However, the user has already provided a project name, "Opening a Restaurant in Tang Dynasty," which suggests a theme of time travel to the Tang Dynasty and running a restaurant. This should be a light-hearted, food-themed novel.

But the user said, "Please help me through dialogue," so I should first engage in conversation with the user to understand their ideas rather than making all the decisions for them.

Wait, let me re-examine this. The user said, "Guide me step by step," so I can start by asking questions to understand the user's preferences and then build based on their responses.

Since the user hasn't provided more information, I can start by asking some questions to understand their preferences.

Considering the project name "Opening a Restaurant in Tang Dynasty" and the pen name are already determined, the user might already have some ideas.

Let me check if there are any skill files that need to be read.

The system prompt says: When the conversation gets too long, the user may start a new conversation or summarize the current one. To use a specific skill, read the skill file.

It mentions the "Plot Tree Construction" skill, stating, "This skill must be used during the initial construction of the novel, the opening, and when building each chapter."

So, I should first read this skill file.

However, I want to first converse with the user to understand their thoughts and then proceed step by step.

Let me see what's currently in the project and then start the conversation with the user.

Actually, the user's message is already included in <user_query>, so I need to respond directly to the user.

Let me first read the skill file to understand the content of the Plot Tree Construction skill.

Okay, I'll read the skill file first and then start the conversation with the user.

It seems the Agent considered all the aspects we set. Let's now look at the results of following the skill. I won't show the intermediate steps, but the final result shows that the skill requirements were generally followed well, though some details were missed. Overall, it's acceptable but still has room for improvement.

For example, the format for creating the synopsis was correct:

However, the last few suggestions in the skill were overlooked:

In total, we had the Agent build a 100-chapter outline based on the Plot Tree Construction skill. The overall plan was for five volumes, but within these 100 chapters, it included content that was supposed to be spread across the five volumes, deviating from the initial plan. However, if we look at just these 100 chapters, the design is logical and meets expectations.

In summary, DeepSeek-V4-Flash scores 80 points for word count control and skill adherence, with 20 points lost on some details. For memory, it scores 70 points, passing but with significant room for improvement. In terms of content creation, without any skills, it might only score 60 points, especially in handling chapter endings, which are somewhat abrupt, and the suspense and hooks are not well designed, requiring skill guidance.

It is recommended that if you use DeepSeek V4, try the Pro version with thinking mode, which should perform better. The Flash version still requires skill or prompt guidance and cannot be fully relied upon. However, based on recent community feedback, V4 Pro does not show a significant improvement in writing capabilities. It's not realistic to expect AI to generate excellent content entirely on its own in the short term. Nonetheless, with appropriate skills and prompts, AI can efficiently complete related tasks. We will continue to evaluate and provide more useful skills to guide DeepSeek V4 in creating better content.