Installed Codex But Don't Know What to Build? Find Real Needs with the "Friction Point" Mindset

It Started with a Comment I Couldn't Read

Many people, after installing AI coding tools like Codex or Claude Code, fall into an awkward situation: they've gotten the demos running, they're comfortable with the tools, but every time they open them, one thought fills their mind—what exactly should I have it build?

Tool Background: Codex is a code generation model from OpenAI, built on the GPT architecture and fine-tuned for programming tasks. It can understand natural language and output runnable code. Claude Code is a command-line AI coding assistant from Anthropic, deeply integrated with the terminal environment, capable of directly reading and writing files and executing commands. These two tools represent the new paradigm of "Agentic Coding"—AI is no longer just a code completion tool, but an executor that can complete entire development tasks end-to-end. Precisely because of this, they have an extremely low barrier to entry, yet they also leave many users in a vacuum between "being able to do it" and "knowing what to do."

The experience of one Bilibili content creator is quite representative. While watching tool reviews on YouTube, he was tormented for a full year by the tiny, fixed-size font in the comments section: his face pressed within 20 centimeters of the screen, straining his eyes to read a top comment three times over. And when a comment contained slang, even with his level-six English certification he couldn't understand it—so he'd copy it, switch tabs, paste it into Google Translate, going back and forth four times for a single comment.

That night he opened Codex, and within half an hour he had a browser extension: the comment font size was now adjustable, and translations appeared directly beneath the comments. The problem was solved, but the truly valuable insight wasn't in the extension itself—it was in how to figure out what to build.

Describing the Need: Skip the Tech Jargon, Just Explain the Symptom

The prompt for this extension was remarkably plain. The creator read it out verbatim:

I've noticed that YouTube's default comment font is too small. Please give me an extension that lets me separately adjust the font size of the comment area, and that can call Google Translate to translate the corresponding comments.

No technical terms, no specified frontend framework. Solving this kind of everyday minor problem doesn't require a framework at all. The only thing to note is to select the Build Web App mode—a browser extension is essentially the same set of code that runs inside a web page.

How Browser Extensions Work: Chrome extensions are built on the WebExtensions API, and essentially consist of HTML, CSS, and JavaScript, declaring permissions and entry points through manifest.json. Their core capabilities include: content scripts (injected into target web pages to manipulate the DOM), background service workers (handling backend logic), and popup pages (the interface that appears when you click the icon). Precisely because an extension is "the same set of code that runs inside a web page," it's entirely reasonable to select Build Web App mode when generating with AI—the output can be loaded and run directly without additional compilation or environment configuration.

Hidden here is the core philosophy of Vibe Coding: you don't need to understand the technology—you just need to describe the symptom clearly. AI doesn't write code perfectly on the first try, and the advantage is exactly that you only need to describe the symptom; it will iterate and revise on its own.

What Is Vibe Coding: This concept was formally introduced by AI researcher Andrej Karpathy in early 2025. Its core proposition is: developers don't need to dig into the implementation details of every line of code—they just describe their intent in natural language, let AI handle the actual writing and debugging, while humans steer the direction and validate the results. The rise of this idea is essentially a redrawing of the "human-machine collaboration boundary" now that AI coding capabilities have matured—the value of humans shifts from "writing correct code" to "asking the right questions." It's not a regression in programming ability, but a democratization of the entry point to programming.

Deployment Is Surprisingly Simple

After generating the code, open Chrome's extensions page, turn on developer mode in the top-right corner, drag the entire code folder in, and the extension is immediately usable. Refresh YouTube, and a new icon appears in the top-right—no registration, no payment, no store listing required. This kind of "write-and-use" personal tool is one of the most underestimated application scenarios in the era of AI coding.

Where Do Ideas Come From? The Answer Lies in "Friction Points"

The more smoothly you use a tool, the easier it is to fall into the trap of "holding a hammer, looking for nails." The creator points out: project ideas have only two sources.

The First Source: Your Own Friction Points

Note that what he emphasizes is not "needs," but friction points—those actions you do every day, do awkwardly every day, and after the awkwardness, silently endure.

The Origin of the Friction Point Concept: The concept of "friction point" comes from the field of user experience design, referring to the resistance or inconvenience a user encounters in the process of achieving a goal. The "problem interviews" in lean startup methodology and Y Combinator's "make something people want" are classic frameworks that, at their core, are all about finding high-frequency, real friction points. In the past, such small friction points were long ignored because "the cost of solving them was too high"—hiring a developer was too expensive, and learning the skills yourself was too slow. But now that AI tools have compressed the execution cost to a matter of hours, problems that were once "long-tail pain points" can, for the first time, be systematically solved.

For example, enlarging the font and copying text to translate—he endured these two things for a full year. Why endure them? Because they were too small, small enough that it wasn't worth searching for dedicated software to solve them. In the past, such needs were swallowed exactly this way: the cost of solving it was far higher than the cost of enduring it.

But now, AI can knock it out in half an hour. This means we need to change a mental habit: don't think "what product can I make," but instead record "what did I put up with today."

He offered a very actionable method: open a checklist in your notes app and jot things down for a week. You'll find you're never short on ideas—it's just that the ideas were never written down. This counterintuitive conclusion is worth remembering: once AI has lowered the cost of execution, what's scarce is no longer ability, but "recorded, real pain points."

The Second Source: Other People's Friction Points

A personal tool only solves your own problem. It's within other people's friction points that the opportunities to package and monetize are hidden.

Where do you find them? Look at where others complain—Xiaohongshu, various forums, the review sections of app stores. Focus especially on negative reviews. That line in a negative review, "if only it could...," is itself a ready-made requirements document.

The Methodological Value of Mining Negative Reviews: Mining negative app-store reviews is a common competitive analysis technique used by Silicon Valley product teams. Review data from the App Store and Google Play can be scraped and analyzed in bulk with tools like Sensor Tower and AppFollow, but for individual developers, simply reading negative reviews is just as effective. The core value of negative reviews is that users have already completed the first two steps—"discovering the problem → clearly articulating it"—so the developer only needs to complete the third step: implementing the solution. Combined with AI's code generation capabilities, the entire cycle from finding a negative review to getting a working prototype demo can theoretically be compressed to within a few hours—something that was completely unimaginable in the past.

The method is also straightforward: copy the whole thing to the AI, have it first propose solution approaches, then pick one direction and write it out. The creator calls this step the "watershed"—crossing over from solving your own problem to solving the problem of a whole group of people.

Summary: Where the Real Bottleneck in AI Coding Lies

On the surface, this case is about teaching people how to build a browser extension. In reality, it reveals a deeper truth:

When AI can already write code for you, the bottleneck shifts from "how to do it" to "what to do." Once the technical barrier is flattened, the people who can continuously produce value are those willing to notice the subtle frictions of daily life and write them down.

• Your own friction points → Craft dedicated personal tools to boost daily efficiency
• Other people's friction points → Mine negative reviews for scalable product opportunities

For everyone who has installed Codex but doesn't know what to build, this may be the most practical lesson: before rushing to write code, spend a week collecting the small things you've been putting up with.

Key Takeaways