Language

让 Agent 自己造插件 Let the agent build its own plugins

直接告诉它你想要的工作流是什么样的，想造什么类型的插件都可以直接和它说。以下是各类型插件的提示词示例，可以替换为自己真实的需求。 Just tell it what workflow you have in mind — whatever kind of plugin you want, say it in plain words and it will build it. Below are example prompts for each plugin type — swap in your own real needs.

界面插件Interface plugin

学习工作台A study workbench

UI Plugin
area: panel

上传课件，让它分阶段讲、出题、批改作业，所有进度都落在界面上。 Upload the course material, get staged lessons, practice sets, and graded homework — with everything visible on screen.

提示词Prompt

请你帮我设计一个学习工作台ui插件，希望可以让我上传文件，然后会发消息给你给我分多阶段讲解这门课程，还要为我编写带答案的练习题和不带答案的作业，我可以上传作业，会发消息给你让你帮我批改作业，希望我能在ui界面上看到所有信息

Please design a study workbench UI plugin for me. I want to upload files, then have you walk me through the course in stages; write practice questions with answers and homework without answers; let me upload my homework so you can grade it; and see all of that information on the UI.

感官插件Sensor plugin

飞书消息桥A Feishu message bridge

Sensor
Python

在飞书里跟机器人说话，等于在会话里跟 PurrCat 说话。 Talking to the bot in Feishu becomes talking to PurrCat in a session.

提示词Prompt

帮我写一个飞书 sensor，让我能在飞书里的智能体应用里和你直接对话。

Write me a Feishu sensor so I can talk to you directly from the bot app in Feishu.

编排插件Graph plugin

股票分析流水线A stock analysis pipeline

Graph
dashboard

输入一只股票，先收集信息，再分情绪面、技术面、基本面三路分析，汇成看好与看空两份报告，最后交决策人审核。 Input a stock, gather the information, analyze it across sentiment, technicals, and fundamentals, merge into bull and bear reports, and have a decision-maker review.

提示词Prompt

帮我设计一个股票分析graph，需求是输入目标股票，能先收集足够信息后，分情绪面、技术面、基本面三路分析，最后汇总分看好和看空两路，将两个报告汇总到决策人上，再人工审核。

Build me a stock analysis pipeline: input a target stock, gather enough information first, then analyze it along three branches — sentiment, technicals, and fundamentals — merge the results into a bullish and a bearish report, hand both to a decision-maker, and then have a human review.

主循环插件AgentLoop plugin

自媒体选题助手A media topic assistant

AgentLoop
trigger: request

收到选题需求先读沙盒里的账号画像，再抓当日热搜，结束前用 Memo 归档记忆。 On a topic request, read the account profile in the sandbox, pull today's hot searches, and archive the memory with Memo before the loop ends.

提示词Prompt

帮我写一个自媒体运营用的 agentloop，在收到有关选题的需求后会提醒 agent 先去沙盒里的 media/profile.md 里看看账号画像，再调用工具获取对应平台的当日热搜，优先考虑用特定的 MCP 工具，没有工具时使用 Search 和 Fetch 工具。在循环结束前会检查有没有用 Memo 工具归档记忆。

Write me an agent loop for running a media account: on receiving a topic request it reminds the agent to first check the account profile in media/profile.md in the sandbox, then calls tools to fetch that platform's hot searches for the day — preferring a dedicated MCP tool, and falling back to Search and Fetch when there is none. Before the loop ends it checks whether the memory has been archived with the Memo tool.

技能Skill

技能开发推荐从一次成功的轨迹开始，先和 agent 在没有 skill 或使用已有的 skill 的情况下完成这个任务后，再使用 trace2skill 功能将经验积淀为新 skill 或改进已有 skill。此部分也可以将你以往有的 skill 直接导入，确保符合通用规范即可无缝对接已有生态，以下是 trace2skill 的提示词示例： Skill development works best starting from a successful run: first finish the task with the agent using no skill, or an existing one, then use the trace2skill feature to distil that experience into a new skill or improve an existing one. You can also import skills you already have — as long as they follow the common spec, they plug into the existing ecosystem seamlessly. Here is an example of a trace2skill prompt:

把团队规范做成技能包 Turn your house style into a skill

Skill

语气、术语表、引用格式一次写清楚，以后每次写作都照办。 Tone, glossary, and citation format written once, applied every time.

提示词Prompt

请你将这次定稿的文章的风格固化为一套规则，写进一个新的 skill 里面，就叫 my-writing-style-skill。

Distil the style of the article we just finalised into a set of rules and write them into a new skill, named my-writing-style-skill.

MCP 工具MCP tools

可将已有的 MCP 工具的配置直接写入配置文件里，直接可以生效，如有自己想实现的特殊工具，也可以和 Agent 慢慢探索，以下是开发一个可供 agent 调用的免费生图工具示例： You can write the configuration of an MCP tool you already have straight into the config file and it takes effect immediately. For special tools you want to build yourself, you can also explore them step by step with the Agent. Here is an example of building a free image-generation tool the agent can call:

造一个免费生图工具Build a free image tool

让 agent 自己设计一个调用免费生图模型的 MCP 工具，边试边调。 Have the agent design an MCP tool that calls a free image model, iterating as it goes.

提示词Prompt

请你设计一个能让你调用网络上的免费生图模型来生图的 MCP 工具。

Design an MCP tool that lets you call a free image-generation model on the web to generate images.