InertiaRSS Track and read blogs, news, and tech you care about
Read Original Open in InertiaRSS

Recommended Feeds

博客园 - 司徒正美
V
V2EX
T
Tailwind CSS Blog
有赞技术团队
有赞技术团队
aimingoo的专栏
aimingoo的专栏
Apple Machine Learning Research
Apple Machine Learning Research
IT之家
IT之家
Blog — PlanetScale
Blog — PlanetScale
A
About on SuperTechFans
月光博客
月光博客
T
The Blog of Author Tim Ferriss
宝玉的分享
宝玉的分享
Martin Fowler
Martin Fowler
博客园 - 聂微东
The GitHub Blog
The GitHub Blog
V
Visual Studio Blog
WordPress大学
WordPress大学
酷 壳 – CoolShell
酷 壳 – CoolShell
Engineering at Meta
Engineering at Meta
GbyAI
GbyAI

阮一峰的网络日志

科技爱好者周刊(第 396 期):互联网通信的替代方案 科技爱好者周刊(第 396 期):互联网通信的替代方案 - 阮一峰的网络日志 科技爱好者周刊(第 395 期):软件开发的第三种方式 科技爱好者周刊(第 395 期):软件开发的第三种方式 - 阮一峰的网络日志 科技爱好者周刊(第 393 期):脑腐状态 科技爱好者周刊(第 392 期):axios 投毒与好莱坞式骗术 科技爱好者周刊(第 391 期):AI 的贫富分化 科技爱好者周刊(第 390 期):没有语料,大模型就是智障 套壳中国大模型撑起500亿美元估值?扒一扒 Cursor 的"套壳"疑云 科技爱好者周刊(第 389 期):未来如何招聘程序员 科技爱好者周刊(第 388 期):测试是新的护城河 零安装的"云养虾":ArkClaw 使用指南 科技爱好者周刊(第 387 期):你是领先的 科技爱好者周刊(第 386 期):当外卖员接入 AI 字节全家桶 Seed 2.0 + TRAE 玩转 Skill 科技爱好者周刊(第 385 期):马斯克害怕中国车企吗? 智谱旗舰 GLM-5 实测:对比 Opus 4.6 和 GPT-5.3-Codex 科技爱好者周刊(第 384 期):为什么软件股下跌 科技爱好者周刊(第 383 期):你是第几级 AI 编程 Kimi 的一体化,Manus 的分层 科技爱好者周刊(第 382 期):独立软件的黄昏 AI native Workspace 也许是智能体的下一阶段 科技爱好者周刊(第 381 期):中国 AI 大模型领导者在想什么 科技爱好者周刊(第 380 期):为什么人们拥抱"不对称收益" 科技爱好者周刊(第 379 期):《硅谷钢铁侠》摘录 我如何用 AI 处理历史遗留代码:MiniMax M2.1 升级体验 科技爱好者周刊(第 378 期):预测是新的互联网热点 科技爱好者周刊(第 377 期):14万美元的贫困线 科技爱好者周刊(第 376 期):太空数据中心的争议 科技爱好者周刊(第 375 期):一扇门的 Bug 终于有人做了 Subagent,TRAE 国内版 SOLO 模式来了 科技爱好者周刊(第 374 期):6GHz 的问题 VS Code 使用国产大模型 MiniMax M2 教程 科技爱好者周刊(第 373 期):数据模型是新产品的核心 国产大模型接入 Claude Code 教程:以 Doubao-Seed-Code 为例 科技爱好者周刊(第 372 期):软件界面如何设计 大模型比拼:MiniMax M2 vs GLM 4.6 vs Claude Sonnet 4.5 科技爱好者周刊(第 371 期):一个乐观主义者的专访 科技爱好者周刊(第 370 期):正确的代码高亮 错误处理:异常好于状态码 科技爱好者周刊(第 369 期):Tim 与罗永浩的对谈 科技爱好者周刊(第 368 期):不要这样管理软件团队 一天之内,智谱和 Anthropic 都发了最强编程模型 科技爱好者周刊(第 367 期):Nano Banana 的几个妙用 科技爱好者周刊(第 366 期):旧金山疯狂的 AI 广告 科技爱好者周刊(第 365 期):流量变现正在崩塌 科技爱好者周刊(第 364 期):最难还原的魔方 科技爱好者周刊(第 363 期):最好懂的神经网络解释 科技爱好者周刊(第 362 期):GitHub 工程师谈系统设计 科技爱好者周刊(第 361 期):暗网 Tor 安全吗?
Shortcuts for AI Development: Workflow Mode
阮一峰 · 2024-10-31 · via 阮一峰的网络日志

I. Introduction

Most people using AI are probably at a beginner level, just like me.

On a daily basis, they ask AI questions (also known as chatting) and occasionally use some existing services: generating images, writing code, translating articles, and so on.

However, over time, they find it insufficient. Many personalized needs cannot be met with existing tools, and they have to figure things out themselves.I think this is a problem many people face:if you want to use AI deeply, you must develop it yourself.

Today, I will introduce a simple method for AI development that doesn't require high门槛, and ordinary people can also get started.

II. Three Modes of AI Development

Following the difficulty level, I divide AI development into three modes.

(1) Self-built models:Build AI models yourself and find data to train them.

This model has high difficulty and cost, and it must be used by AI professionals or large companies.

(2) API Call : Write your own scripts to call the API of models run by others.

This model requires users to be proficient in programming. The advantage is strong adaptability, as any demand can be solved by programming. The disadvantage is that writing and debugging scripts takes a lot of time, and you need to deploy a server yourself.

(3) Workflow Model : Use workflows provided by AI service providers to orchestrate each step, forming automated operations.

This model is the simplest, and even those who don't understand programming can get started. It doesn't require writing scripts, and there are ready-made servers, saving time and effort. The disadvantage is that it is limited by the capabilities provided by the AI service provider; if they don't offer it, you can't do it.

I'm a beginner in AI, so I chose the last mode and tried AI development.

III. GPT Store and Copilot Templates

Many major AI companies offer "workflow modes," allowing developers to customize their own applications on their models.

The most famous of these is probably the GPT Store from OpenAI.

It opens up the underlying capabilities of ChatGPT, enabling developers to customize specialized GPTs for specific fields, such as academic paper GPTs, Excel file processing GPTs, and so on.

Microsoft's Copilot also has a similar feature called "template."

However, these are all foreign services that require payment, and the servers are located overseas. Fortunately, there are free alternatives in China, so I chose a domestic product.

IV. Workflow Demonstration

I'll first demonstrate the workflow's operation interface. Everyone will then understand what it is.

A workflow is essentially breaking down an AI application into several steps . Each step is a node, where you set the operation for this step, and all nodes complete the task according to the process.

The image above is the workflow编排界面 of 扣子 AI (coze.cn), which includes four nodes.

扣子 is a free AI application development platform under ByteDance. The process is very intuitive—just use a web canvas to create AI applications step by step through graphical operations. I used it for the demonstration.

Its development page on the left side shows different types of nodes, and the following image shows part of them.

When creating a new node, just click the corresponding type. Generally, the first step is always to create a "model node" or "plugin node".

After creating a model node, if you want to train the model with your own data, you can create a knowledge base on the project management page and upload the data file. Then, within the workflow, create a "knowledge base node".

If you want to enable persistent memory to remember users' historical conversations, you can create a database to save user data, and then create a "database node" in the workflow.

Once all the nodes are set up, this AI application can be run. The final step is to publish the application.

The button will prompt you to select the publishing channel (as shown in the image above). Currently, it supports publishing to stores (using the Kedu platform), third-party platforms (using Lark, Douyin, WeChat), and APIs and SDKs (for use in users' own software).

is the user interface for the workflow mode, just a simple introduction, so everyone just needs to have a general understanding. For specific operations, you can refer to the document .

It's still very easy to get started with. Once you're proficient, developing an AI application can be done in just one or two hours.

V. Examples of workflow development

To deepen the understanding, I'll give an example of actual development. A catering company wants to build an AI application to analyze user reviews.

For this, create a new workflow with three basic nodes——input node, AI model node for analyzing user reviews, and output node.

Then, publish this workflow as an API. When user reviews are input into this API, it will return a JSON data segment containing the user reviews identified by the AI model (positive, negative, neutral), as well as review keywords (environment/service/dishes).

The above is a preview and debugging interface of this workflow. An actual review was input, and the running result was returned in JSON format.

It can be seen that the AI identified the user's review as a neutral rating for the product, a negative rating for the environment, and a positive rating for the service speed.

VI. Template Functionality

In addition to personal use, the workflow can be saved as a template (template) and published to the Template Store for others to use and reference.

In the Template Store, you can see workflows built by the official team and other users. Currently, there are already thousands of them, and the number is still growing rapidly.

These workflows can be used directly or copied as templates. After copying the templates, users can learn from and modify them to adapt them to their own applications.

For template authors, they can set a price for the templates to achieve monetization.

Many templates offer very interesting features, worth checking out. For example,Douyin to Xiaohongshu".template.

You input a short video link from Douyin, and AI automatically generates 10 usable titles and a copy. You can then post them along with the video to Xiaohongshu.

7. Tutorial: Expert in Emotional Story Viral Content

There are already many tutorials for developing workflow templates online. You can search for them.

I saw an article titled "Expert in Emotional Story Viral Content"Template tutorialIt's very interesting. Emotional web novels are currently very popular in terms of traffic, so the author developed an AI application specifically for generating such articles.

First look at the finished product, clickThe template link,try it out. You input an article topic, like

a couple meeting in university, breaking up after graduation, and meeting again 10 years later.

This AI app will output the outline and main text of the article.

It also includes images, which can be published directly.

Its underlying workflow is roughly divided into six nodes.

  1. Start node: receives the user's input article topic
  2. AI model node: generates the article outline based on the topic using an AI model.
  3. Expansion node: expands the outline to generate the main text.
  4. Content summary node: generates a summary of the article based on the main text.
  5. Text-to-image node: generates images for the article based on the summary.
  6. End node: Summarizes all content and outputs the result.

After setup is complete, click the "Publish" button, and this AI application is ready to use.

Section VIII: Conclusion

The workflow mode is much easier and faster compared to writing from scratch, and it also solves the server-side issues. If you have AI development needs, you can try this mode.

Next, why not start from the documentation to create your first workflow node and publish it to the template store.

For any development issues (which you're likely to encounter), you can join the official consultation group to ask. Here is the WeChat group QR code.

When I encountered a problem, I consulted a classmate named Koushi. They told me that there are currently two campaigns with plenty of prizes, and recommended that I participate.

I think it's great. I'll forward the event information below. If you've just made a template, don't miss out.

(1)Template Competition

After the workflow for the template is completed, it will be listed in the template store. You can nowearn rewards.

You can set a price for the template (after official review), to monetize it. Others will need to pay to copy the template. During the event, the official will issue prizes ranging from 1,000 to 10,000 yuan based on the number of template copies, as well as randomly selected 50 wireless chargers as commemorative prizes.

(2)Halloween Giveaway

The official now has aHalloween Giveaway,where you upload your photos, you will get a Halloween-themed photo, as well as a chance to win.

The prizes for the draw include iPhone 16 Pro Max, Switch, and a coffee machine.

(Finished)