AI Hackathon Journey: Building the "Novel Jump" Bot on Coze Platform

Background

On March 30, the author and teammate ETERNCHAN attended the ByteDance Juejin x Coze offline Hackathon in Shenzhen, an AI‑focused developer event aimed at turning ideas into practical applications through hands‑on experimentation.

The hackathon served as a laboratory for converting creative concepts into functional AI solutions.

About Coze

Coze: ByteDance's AI Bot development platform for creating personalized intelligent agents.

Coze integrates plugins, memory, and workflow capabilities, allowing low‑threshold, rapid construction of custom or commercial AI agents that can be deployed to Feishu, WeChat Official Accounts, Juejin, and other channels.

Beyond conversational bots, Coze supports complex AI tasks such as retrieval‑augmented generation (RAG) systems and intricate pipelines.

Novel Jump Concept

The team brainstormed several ideas and settled on an immersive dialogue with novel characters, offering interactive role‑play, scenario simulation, and a potential subscription‑based business model.

The bot, named "Novel Jump" (originally "Novel Jumping"), aims to let users converse with literary characters, bringing stories to life.

Main Features

The development focused on the following functionalities:

Display novel introductions and character lists.

Generate character portraits via text‑to‑image.

Retrieve classic lines and personality traits from a knowledge base.

Enable RAG‑based Q&A with characters.

Development & Debugging

The biggest challenge was the bot's poor instruction‑following ability, especially when prompting for character portraits or maintaining distinct dialogue styles.

To mitigate this, the team introduced workflows to orchestrate the bot's behavior, improving output stability.

The technical stack included:

The workflow architecture consists of three main pipelines:

get_novel_introduction : Retrieves novel summary, character list, and generates a cover image.

get_character_lines_portrait : Fetches classic lines and creates a character portrait.

ask_character : Performs RAG‑based answering for user queries.

Each workflow contains a start node, a knowledge‑base node for semantic retrieval, a large‑model node for processing, and, where needed, a code node for data formatting and a text‑to‑image node for visual output.

Results

"The conversation feels like drawing cards—full of ups and downs, finally achieving that satisfying taste."

After iterative tuning, the bot reached a satisfactory level, though further improvements are planned.

Issues Encountered

"Failed draw!"

Despite using the Moonshot model and workflows, occasional instruction‑non‑compliance persisted, such as the bot revealing internal workflow steps in its responses.

Deployment

The bot was published to the Coze Bot Store, Juejin community, and Doubao, currently running on the Yunque model for public access, with a limited set of available characters.

Coze Bot Store: https://www.coze.cn/store/bot/7353452051638255628

Juejin AI Chatroom: https://juejin.cn/pin/club/7173895829281833012?robotId=7351428879468806144

Awards

The project received five awards, including Best Creativity, Best Technology, Best User Experience, Best Solution, and the overall Champion award.

Atmosphere

The event atmosphere was lively, with participants sharing experiences and enjoying the hackathon spirit.

Conclusion & Reflections

"Technical honing, a leap of thought and creativity."

The hackathon was both a technical challenge and a creative breakthrough, highlighting the importance of combining LUI (Language User Interface) with GUI for better user guidance, improving LLM instruction compliance, and leveraging integrated AI agent architectures.

References

Juejin x Coze Hands‑On Lab – Shenzhen

Andrew Ng’s tweet on LLM‑based agents: https://twitter.com/AndrewYNg/status/1770897666702233815