The Most Detailed Codex Guide on Bilibili: From Installation to Real-World Practice

In an era where AI tools keep emerging, most products are still stuck at the "conversation" level, while Codex has directly crossed that threshold—it can actually get to work on your computer. A recent video on Bilibili titled "The Most Detailed Codex App Guide on All of Bilibili" demonstrates the complete journey of this tool from installation to real-world application using a large number of real projects. Many viewers have called it a must-watch introduction that lets you "master Codex in one episode."

The Essential Difference Between Codex and Doubao, DeepSeek

The video opens by pointing out a key difference: tools like Doubao and DeepSeek^Ad can only chat with you, while Codex can directly operate your computer. Organizing files, creating reports, reading work content, writing weekly reports, making PPTs, operating browsers, building complex software—Codex can complete all of these tasks locally.

What's even more practical is its automation capability: after you set up a task, it can run automatically even while you're asleep, and by the time you wake up, the task is already done. This means Codex's positioning is not a "Q&A assistant" but a "digital employee."

The key to understanding the fundamental difference between Codex and Doubao, DeepSeek lies in the distinction between "conversational AI" and "AI Agent." Doubao and DeepSeek are essentially content generators driven by Large Language Models (LLMs)—you input a prompt, it outputs a text reply, and all capabilities remain at the level of "speaking." Codex, on the other hand, is an AI Agent, which adds two core capabilities on top of the LLM: "tool calling" and "task looping." The former allows it to read and write files, execute commands, and operate browsers; the latter allows it to autonomously decide what to do next based on execution results until the task is complete.

Put simply, if Doubao is a "consultant," then Codex is an "employee" who can personally step onto the field. This shift from "speaking" to "doing" is a key watershed in the evolution of AI products today. Only by understanding this layer can you grasp Codex's true positioning—it's not a smarter chatbot, but a general-purpose executor that can do work on your behalf.

Installation and Basic Configuration

Codex is already available on the Microsoft and Apple app stores—just search for it and install it directly. The interface after opening is similar to the web version of Doubao: an input box in the middle and a sidebar on the left. However, the plugins, automation, workflows, and other features in the sidebar are things Doubao doesn't have.

For first-time use, it's recommended to adjust one configuration first: the work mode. Switch it to "Daily Work" mode—Codex can still complete all tasks, but it won't display a large amount of technical details, making the experience much smoother. The video author emphasizes that Codex doesn't require users to specifically learn technical knowledge—a few sentences can produce cool results. But before that, understanding the concept of "projects" will greatly improve your efficiency.

Projects and Folders: Letting Codex Know Where to Work

Since Codex operates directly on your computer, it needs a working folder—Codex calls these folders "projects," which are displayed in the sidebar. It's recommended to categorize folders that need Codex's processing into different projects by different matters, so you can find the corresponding content at a glance.

After selecting the project for this conversation below the input box, Codex can directly read all the files inside. In the video, the author prepared a real folder called "Computers and Coffee Shop," stuffed with sales records, invoices, inventory lists, and other data, covering file formats including spreadsheets, PDFs, and images.

The concept of "projects" is actually standard practice in the AI Agent field, corresponding to a Workspace or Context Boundary. The context window of a large language model is limited—if you point it at the entire hard drive, it can neither retrieve efficiently nor avoid accidentally operating on irrelevant files. Declaring a folder as a project is equivalent to drawing a clear "work scope" for the Agent—it only needs to focus on files within that scope, which both improves processing speed and reduces the risk of accidental deletion or modification.

In practice, it's recommended to categorize by the principle of "one matter = one project," such as finance, copywriting, and client materials each being independent projects. This way, each conversation only requires switching projects, and the Agent automatically gains all the context of that domain without repeated file uploads or explanations. This habit is crucial for improving long-term efficiency and is a key step from "casually using" to "truly mastering" Codex.

Real-World Task 1: File Organization and Data Extraction

The first task the author sent was: organize the messy coffee shop files and extract and organize the information into an Excel file.

This seemingly beginner-level task is actually quite labor-intensive to execute. Codex needs to analyze files, move files, modify content, and create new files. The author deliberately planted some traps in the files to see if Codex could discover them. A few minutes later, the task was complete: the folder was organized neatly by purpose, sales data and product rankings were categorized into tables, and details from PDF invoices were also extracted.

More importantly, Codex did discover the planted issues—it flagged missing transaction records and pointed out anomalies in two invoices. This proves its level of meticulousness in data processing. At this point, Codex can already easily handle daily tasks like organizing reports, writing documents, and generating daily summaries.

Real-World Task 2: Plan Mode and Visual Dashboard

To test a more complex scenario, the author used the command line to add the summary table just generated, asking Codex to create a visual dashboard displaying sales, expenses, costs, inventory, and other operational conditions.

Facing a complex task, the author switched to "Plan Mode." In this mode, Codex doesn't execute directly but first plans out a detailed execution checklist, confirms details with you, and then executes. The author recommends: relatively complex tasks should all have Plan Mode enabled.

During the waiting process, the author demonstrated an interesting feature—typing a slash in the input box summons a desktop pet. Besides being cute, it also displays all currently executing tasks, and clicking on one jumps to the corresponding task. In Plan Mode, Codex proactively asked the author several questions, and only after the author made selections did it continue working. It ultimately generated a detailed plan and only began actual execution after confirmation.

The dashboard ultimately presented a professional standard, including all requested data dimensions, with suggestions and key items listed below. The author candidly admitted not being a sales professional and unable to judge its practicality, but if you are a practitioner, you can evaluate whether this dashboard can help the company operate better. Subsequently, the author asked to change the theme to "more modern, intuitive, like Apple's style," and Codex completed it smoothly—the result indeed had an Apple style.

"Plan Mode" corresponds to the Plan-and-Execute framework in AI Agent design. The traditional immediate execution mode tends to "make mistakes while doing" when facing complex tasks—once the direction is off, the rework cost is high. Plan Mode forces the Agent to first produce a step-by-step execution checklist, hand it over for human review or supplementary information, and then execute in sequence after confirmation.

This "Human-in-the-loop" mechanism has two benefits: first, it allows users to correct direction at a low-cost stage, preventing AI from running wild based on wrong assumptions; second, it makes the progress of complex tasks predictable and traceable. For tasks involving multiple files, multiple steps, and cross-tool operations (such as from data extraction to dashboard development), Plan Mode is almost mandatory. Using immediate mode for speed on simple tasks and Plan Mode for stability on complex tasks is the basic rhythm of using Codex.

Summary of Usage Recommendations

From the demonstrations in this video, several practical points can be distilled: First, configure the work mode first—Daily Work mode filters out unnecessary technical details; second, make good use of project categorization to give Codex a clear work scope; third, execute simple tasks directly and switch to Plan Mode for complex tasks; fourth, Codex's capability boundary depends on your imagination—from file organization to dashboard development, it can cover it all. For people who don't want to learn programming but want to use AI to improve work efficiency, Codex provides a low-barrier path.