What is this article talking about?
Newbies often understand all AI as a chat box, but chat AI, single web tools, and desktop work agents actually face completely different tasks. This article uses my own experience of changing tools all the way from 2022 to clarify the boundaries of the three entrances, and also explains why the desktop version is a paradigm shift. Who is
· People who have used ChatGPT or Claude to chat, but don’t know what Agent and desktop versions are popular about
· Liberal arts students, clerical workers, teaching and content workers who do not write programs
· People who have accumulated a lot of data in project mode and start to feel stuck
· The division of tasks between the three AI portals, clearly distinguish them before choosing tools
· Three real pain points of the web version project model and the solutions they lead to
· Desktop Agent’s three-layer division of labor workflow, with open source skills package link
First distinguish the three entrances
ChatGPT, Claude, Gemini dialog box. Suitable for discussion, rewriting, summarizing, and brainstorming, and helps you organize your thoughts at the textual level.
Tools such as image generation, presentation generation, and video summary are available right out of the box. It's easy to get started, the task boundaries are clear, and the data always stays on their respective platforms.
A work assistant that can be installed into a computer and can read local folders, modify files, and run processes. The desktop versions of Codex and Claude all fall into this category.
No one of the three entrances can replace the other, just choose based on the task: if you just want to discuss text, the chat box is the fastest; if you want to quickly generate a picture, the web tool is the easiest; if you want to organize a batch of local files, update website information, and complete a process, it is the turn of the desktop Agent.
What the web version era taught me: Project model and its three pain points
Before the desktop version, the first time I really felt the change in workflow was the project mode of ChatGPT. Throw in the reference materials for specific topics, and you can also set up project instructions. What impressed me the most was using it to prepare for the iPAS AI Application Planner exam: I threw the syllabus and teaching materials into the project, and the day before the exam I told it, "organize the things we talked about in the past, and make a page of cheat sheets for me that I am not familiar with, often confused, and keep forgetting." It was done. At that moment I deeply felt: it understands me and remembers what I need, so there is no need to repeat it every time.
After using it for a long time, three pain points emerged honestly:
- The accumulation of errors cannot be deleted: nine wrong questions were asked, one correct question was asked, and wrong knowledge and correct knowledge were all mixed in the same project.
- Cross-project analysis is difficult: I want to combine the key points of three projects A, B, and C into D. There are almost no good methods at the moment I use.
- Moving is very painful: If you want to change a platform, all the information must be reorganized.
(These three points are the actual situation when I used them. The platform functions are constantly being updated. The actual limitations are subject to the version you have. The key point is the direction they point to together.)
The third pain point made me make a decision that affects me now: store the data locally and use my own folder as my home. NotebookLM is also a magic weapon during this period. It is a strong personal knowledge base, and its strength is query and retrieval. I will also select my work data together with a new paper, and ask it, "Based on my work habits, can I use the content of this paper?" Let the knowledge base screen first, and only read the ones that can be used, which reduces knowledge anxiety by half.
Paradigm shift for desktop: AI comes to work in your folder
In the desktop version era, the concept is completely reversed. I open a folder on my desktop that Codex can read, Claude's desktop version can read, and any tool that comes after that can read it. The files are here with me, and all the AIs take them over, and they can divide the work and cooperate. My files are my system.
This path is particularly important for liberal arts students, because there is a common misunderstanding that needs to be dispelled first: People think that AI is most helpful for writing programs because engineers know AI better. In fact, we are using a large language model, and program code is just one of the highly structured human languages. As long as the knowledge is organized and systematized, AI can also empower your interviews, meeting notes and stories.
There is a story circulated within Anthropic: A data management expert in the team went to learn to write programs. His supervisor asked him why he was learning this. He said, "I see everyone is so good at using Claude Code. I also have a lot of files to organize, so I had to learn it." At that moment, they realized that there were a large number of knowledge workers in the world who needed to organize documents. They did not need to become engineers. They also needed in-depth analysis and could not make mistakes easily. A large part of the subsequent development of desktop tools was to serve this group of people.
Codex: GPT desktop version installed on your computer
Desktop Agents such as Codex are most easily understood as the GPT desktop version installed on your computer. It goes out of the web chat box, reads files, writes files, cleans data, batches sorting, and generates web pages in the folder you specify, leaving the work results in your project. Suggestions for newbies: Start with a desktop Agent, which has a lower threshold, to get used to the Agent's working methods and data collection process. Later, if you need in-depth analysis, add a second tool for cross-comparison.
An actual three-layer division of labor workflow
The workflow I run most often now connects three tools into three layers:
Catch sources, generate subtitles, and save data so that the content can be traced and reviewed.
Organize, classify, label, and save into the local knowledge base.
In-depth analysis, comparison, assisted decision-making, and read the cleaned version.
When there is a small amount of data, the inference layer can directly read NotebookLM; for a large amount of data (such as three hundred videos on a channel), the cleaning layer should first organize, label, and save it locally, and then let the inference layer read the compiled version. I made this into an open source skill package notebooklm-connector. After installing it, you can directly check your NotebookLM in the Agent dialog. GitHub can be obtained. A complete demonstration of how this three-tier division of labor actually works can be found inTurn a YouTube into an opinion reporting websiteThat article.
How you can get started
Before using AI next time, ask yourself a question: Do I just want to discuss text, or do I want to organize data, operate files, and complete a process? Different answers mean different entrances.
- If you are still in the chat box stage, first open common topics into projects to experience the difference of "it remembers you" and experience the three pain points personally.
- After the pain point appears, open a local folder as a test site, install a desktop Agent, and let it report what is in the folder first.
- Give it a small task: organize ten notes, change file names in batches, and clean up a verbatim draft into an organizational draft.
- If you have a smooth process, write it down as a skill package and keep it with you. You can take it with you when you change tools later.
If you want Agent to do things, you must make the work clear: which files can be moved and which cannot be moved, whether they need to be deployed after they are completed, and who will check and accept them. This is why Agent teaching is not enough to teach prompt words. What is more important is workflow, knowledge base, file rules and acceptance. Liberal arts students can take this path, and there is no need to write programs.