Building and Optimizing AI Bots with Coze: A Step‑by‑Step Low‑Code Guide

1. Introduction

The author discusses the problem of information overload and the need for tools that can filter and summarize content efficiently. After experimenting with OpenAI GPT‑4 and encountering access hurdles, the author discovers Coze, a Chinese low‑code AI Bot platform that promises bot creation without programming skills.

2. Simple Three‑Step Bot Creation

2.1 Create Bot

Log in to the Coze website using a phone number or Douyin account, click the "Create Bot" button, and fill in the workspace (default is fine), bot name, description, and icon. Confirm to generate the bot and enter the "Orchestration" page.

2.2 Orchestration (Arrange)

The orchestration interface is divided into three panels: "Persona & Reply Logic", "Bot Configuration Options", and "Preview & Debug".

2.2.1 Persona & Reply Logic

Write a prompt (essentially a set of instructions) for the bot. The UI offers an "Optimize" button that lets the AI improve the prompt, and a "Use" button to apply the optimized version.

2.2.2 Bot Configuration Options

Key configurable items include:

Model settings (e.g., "Cloud Sparrow" model with context rounds).

Plugins (e.g., external APIs for search, image generation, etc.).

Workflows (visual composition of nodes).

Knowledge base (upload files or URLs as data sources).

Database (structured table storage).

Opening line and user‑question suggestions.

Voice tone selection.

For a quick test, the author enables only the "Plugin" and "Opening line" options, adding the "LinkReaderPlugin" to extract titles and content from URLs.

2.3 Publish

After filling required fields (publish record and target platform), the bot can be published to Feishu. The author demonstrates the authorization flow, selects Feishu, and confirms publication, after which the bot appears in the Feishu app store.

3. Advanced Usage

3.1 Prompt Optimization

The official documentation recommends using structured Markdown prompts, defining a persona, describing functionality and workflow, specifying output format, and constraining the answer scope.

3.2 Designing a Workflow

A workflow consists of nodes that can be basic (large model, code, knowledge base, selector) or plugins. The default workflow includes a Start node and an End node.

3.2.1 Create Workflow

The author creates a workflow aimed at reliably producing a content summary and a detailed outline.

3.2.2 Add Plugin Node

Search for and add the "LinkReaderPlugin", configure its parameters, and connect it in the flow.

3.2.3 Add Large‑Model Node

Duplicate the plugin node, set a prompt that asks the model to generate a detailed outline using the variable {{input}}.

3.2.4 Add Code Node

Insert a code node (JavaScript or Python) to merge the outputs of the two model nodes into a single response. The code runs in an async main function with an Args input and Output return value.

# 角色
你是一个内容分析专家，擅长对用户输入的 URL 或上传的文件进行分析，包括内容提取、内容分析，还能输出通熟易懂的内容概述和归纳后的详细大纲。

## 技能
### 技能 1: 内容提取
1. 对用户输入的 URL 或上传的文件进行分析，提取其中的关键信息和主要内容。
2. 对提取的内容进行分类和整理，以便更好地理解和分析。

### 技能 2: 内容分析
1. 使用自然语言处理技术对提取的内容进行分析，包括但不限于词汇、语法、语义、情感、关键词等方面。
2. 根据分析结果，生成内容概述和详细大纲，以便用户更好地理解和使用。

### 技能 3: 输出结果
1. 对提取和分析的内容进行总结归纳，生成一个简短的内容概述，让用户快速了解文件或 URL 的主要内容。
2. 根据提取和分析的内容生成一个详细的大纲，包括主要内容、次要内容、细节等，让用户更深入地了解文件或 URL 的内容。
3. 请参考如下格式回复：
**内容概述**：xxx
**详细大纲**：
1.xxx
 - xxx

## 限制
- 不仅要处理文本相关的内容，图片、音频、视频等多媒体内容也要进行处理。
- 输出的内容必须符合事实，不能包含虚假信息。
- 大纲要尽可能详细，尽可能不要错过一点内容。
- 输出结果必须包含两个部分：内容概述 和 详细大纲。
- 内容概述以段落的形式输出，详细大纲以无序列表的形式输出。

3.2.5 Run and Publish Workflow

After connecting all nodes, set the End node to "Use the configured content directly as the answer", run a test, and then publish the workflow for use in the bot.

3.3 Knowledge Base and Database

Coze offers two storage options: a vector‑search enabled knowledge base for unstructured documents, and a relational‑style database that can be queried via natural language. Typical use cases include supplementing corpora, FAQ bots, and vertical domains such as automotive specifications.

3.4 Plugins

Plugins are tool collections that expose APIs to the bot. The platform provides about 60 built‑in plugins, and users can create custom ones (e.g., a WanAndroid API plugin) by configuring icon, name, description, URL, headers, and authentication method.

4. Conclusion

The author reflects that experimenting with Coze’s low‑code bot building features is enjoyable, though some UI details need polishing. They invite readers to join their Coze team, share custom plugins and workflows, and look forward to future API‑based bot invocation.