Clawdbot is booming, and I can't hold back: A zero-background liberal arts student's practical Vibe Coding

Author: Denise I Biteye Content Team, Bbo I XHunt Operations Team

Recently, I saw everyone using Clawdbot on X and leveraging AI agents to do all kinds of tasks. Suddenly, I felt that liberal arts students can also play with code in a “vibe” way, so I started this zero-base practical experiment.

Vibe Coding is regarded as a new paradigm of “requirements as code,” and is even expected to “allow non-programmers to build tools.”

However, a persistent real-world question remains: when the technical threshold is lowered, does the actual implementation ability also come along?

To answer this question, I, a user with no coding experience whatsoever, fully experienced three mainstream Vibe Coding tools.

The process was not smooth: pitfalls, crashes, repeatedly overturning and starting over, but precisely because of this, I identified which parts are most prone to problems and how to avoid these pitfalls.

This article is not a tool review but a genuine record from cognition to practice.

1. Concept Understanding: What is Vibe Coding?

It is a new development model of “just specify requirements, no need to write code.”

This concept was first proposed by former OpenAI co-founder Andrej Karpathy. His words were:

“This is a new way of programming, I call it Vibe Coding. You entrust everything to feelings (Vibes), embrace exponential efficiency growth, and even forget about the existence of code.”

Under the Vibe Coding mode, you are no longer a “construction worker,” but more like a “product manager” or even a “client.”

Your task is to accurately tell AI: “Build a house here with floor-to-ceiling windows, and there a swimming pool.”

If you’re not satisfied, just say: “The vibe is off, let’s change it.”

That’s Vibe!

2. Pitfall Avoidance Guide: The Safety Red Line of Vibe Coding

Before opening the door to AI programming, you must first lock your “treasure vault.” Many beginners unconsciously send sensitive information to AI during conversations, which is extremely dangerous.

1. Three types of data that must never be leaked

• API Key: Equivalent to your “stored-value card” for various platforms. If leaked, the quota can be drained instantly.
• Private keys/mnemonics: Absolutely, forever, do not send to any AI or write into any code files. Once leaked, your on-chain assets will be zeroed out.
• Cookie data: Some tutorials teach you to get cookies from browsers for automation, but cookies contain your login state. If leaked, others can log into your Twitter, Discord, or exchange accounts without a password.

2. Core principle: Environment isolation, data desensitization

Never write API keys or sensitive info directly into code or prompt words.

The correct approach is to use “environment variables,” so the code reads system variables instead of directly exposing keys.

Windows setup steps:

• Press Win+S, type “Environment Variables” in search
• Create new variables
• Enter variable name and value, then save

MacOS setup steps:

• Press Command + Space, search for “Terminal” and open it.
• In the black window that pops up, copy and paste the following command (replace with your real key), then press Enter:

echo 'export CRYPTOHUNT_API_KEY="your_sk-xxxxxx"' >> ~/.zshrc

• To make the configuration take effect immediately, continue with:

source ~/.zshrc

• (Optional) Verify success by typing: echo $CRYPTOHUNT_API_KEY. If your key appears on the screen, the setup is successful.

3. Practical (Crash) Review: Three tools, three experiences

Having finished the theory, let’s jump into practice. To find the most handy tool, I tried three mainstream options. They were chosen because they each represent one of the three main forms of current Vibe Coding:

• Google AI Studio: The “rapid prototype generation” lightweight route
• Antigravity: The “AI agent + full-stack local development” engineering route
• Lovable: The “dialogue as application” abstract route

If you only want to know which tool is most suitable for zero-base direct implementation, you can skip directly to the third tool, Lovable; if you want to avoid detours, I recommend reading this section thoroughly. Now, let’s review in order.

1. Google AI Studio

Positioning: Suitable for individual developers and startups

Experience: Focused on “speed” and “free,” allowing you to verify an AI idea within minutes

Official site: https://aistudio.google.com/apps

My first use was on Christmas. Everyone was making cyber Christmas trees, and I joined in. I casually threw in a prompt, and in just 5 minutes, it generated a very beautiful 3D page with particles flowing with gestures, extremely smooth.

This made me marvel at how much times have changed.

(Think back to the days when tech guys stayed up all night coding to make a heart beat, running for half a day to see the result—feeling pretty cool. But now? You can hand it over to vibe coding, and it can instantly maximize emotional value.)

Buoyed by successful cases, I started to “inflate” myself, then got serious: calling Cryptohunt API to make a Twitter data dashboard.

I confidently fed the API documentation to it. Everything started smoothly, backend logic was instant, data was successfully fetched. I thought: Is that all?

But when I asked it to display the data beautifully on the frontend, the nightmare began:

• Charts not showing? It modifies the code
• Showing but incomplete? It modifies the code again, breaking interactions
• Interaction fixed? Now the API data stops loading

I fell into an endless cycle of “bug fixing—creating new bugs.” After two days, I gave up.

Final display chart (AI already started doodling nonsense)

Pitfall summary: Google AI Studio is a bit of a “specialist.” Great for creative projects or single-page demos. But once it involves deep front-end/back-end integration (like API data flow + complex frontend display), it easily loses focus. If you don’t understand code logic at all, you might get stuck in a dead end on frontend display like I did.

2. Antigravity

Positioning: An integrated multi-model switching proxy-style full-stack development tool

Features: Multi-model switching, task decomposition, automated testing

Official site: https://antigravity.google/

In actual use, I almost gave up before even starting to code because of its login process. Unlike most tools that you can start using immediately after opening the webpage, it has strict requirements on account and network environment.

Account restrictions are not determined by your current IP but by the “associated country/region” of your Google account. If the region isn’t supported (e.g., Mainland China, Hong Kong), login will be rejected. My requests to change regions were also repeatedly denied.

The whole process can only be described as: exhausting.

After three accounts were banned and multiple environment switches, I found a workaround: Antigravity tools

1. First, set a US IP + global VPN mode
2. Download antigravity tools

Shared link from Twitter user: https://x.com/idoubicc/status/2004848130693759213 3. After downloading, add your account and start OAuth authorization

4. Successful authorization redirects to login page, where you can see your model quota directly

(Risk warning: Antigravity Tools is an open-source project designed to connect Antigravity models with tools like Claude Code. During Vibe Coding, it’s recommended to isolate your operation environment from your wallet or important accounts to reduce potential security risks.)

Inside the tool, I also summarized two extremely important tips for beginners:

1. Install Chinese language pack: Search “Chinese” in the extension sidebar, install Simplified Chinese language pack, and restart the software. This makes the interface Chinese, greatly reducing comprehension costs.
2. Inject “Global Persona”: Antigravity allows you to set a “global rule” that the AI must always follow. How: click the Rules icon on the right → select + Workspace (add workspace rule) and paste a “global rule prompt.” It can enforce the use of Chinese and significantly improve code output quality.

However, even after successfully logging in and basic setup, when I started running code, I was met with a cold splash of reality. It can’t preview immediately like Google AI Studio. I had to download additional preview software like Node.js, VS Code, or just view results in HTML. Also, the response speed of large models is visibly slow. Every time I made a request, the screen kept spinning.

But Antigravity has a very impressive feature: automatic testing. After writing code, it launches a video preview, clicks on the webpage like a human, and checks if the logic runs smoothly. Watching AI click around on the screen gives a real sense of “the future is here.”

I tried to make a project: “A cryptocurrency Alpha signal catcher. Use Cryptohunt API to monitor Twitter data, automatically fetch tweets’ Ticks, analyze which tokens are being discussed, and rank by discussion heat.”

Reality was harsh. Despite its effort to help me write code and test, I still hit a wall when connecting to the API. Possibly due to complex local network environment or AI’s misunderstanding of the API documentation, data could not be connected successfully.

Demo: simulated data display effect

Seeing a bunch of red errors in the terminal and AI hallucinating “this time it will work,” I realized: full-stack is great, but due to network and environment black boxes, making a truly usable tool remains very high for beginners. Because Antigravity assumes you’re already a developer.

3. Lovable

Positioning: “Dialogue as Application” platform

Features: No local environment setup needed, what you see is what you get

Official site: https://lovable.dev/

Using Lovable gave me a real taste of being a “client boss.” Is it because only I need to pay? (Tip: I spent 15 RMB on 100 credits on Taobao, and I haven’t used them all yet)

With Lovable, I completed the following crypto Twitter monitoring tool—Crypto Pulse dashboard—without changing a single line of code.

Yes, you read that right, I finally succeeded.

Online experience: https://tweet-whisperer-dash.lovable.app/

[“Feature Details”]

A. Panorama Data & AI Summary

On the homepage, the “thermometer” of content:

• Real-time data cards: Top show total tweets, interactions, active KOLs, market sentiment index within 24 hours. No need to read details; just compare with the previous 24 hours. Glance at red or green to know if the market today is panic or greed.
• AI Smart Summary: The most time-saving feature. Integrated with Gemini AI, it automatically reads thousands of tweets from the past few hours and writes a market brief for you.

B. Trend Radar

Don’t want to miss the next 100x coin?

• Intelligent classification: Automatically filters hot topics into Topic, Project, Token, and KOL.
• Up/down trend indicator: Next to each label, shows the change rate of mentions in 24 hours (up/down). Where discussion surges, opportunities follow.

C. Smart Information Stream

Say goodbye to messy timelines. Here are four cleaned tabs:

• Highlights: High-value content filtered by AI, removing meaningless spam.
• Hot: Tweets with the highest current interactions (likes + retweets).
• Big V: Real-time posts from industry giants (like Elon Musk, Vitalik, etc., with >100K followers).
• Latest: The newest tweets from CT-related accounts, always discovering fresh content.

*Experience details: Supports regional filtering (English/Chinese), clicking on a card jumps directly to the original Twitter post.

[“Usage Reminder”] Since I am using my personal API Key now, data transmission consumes API Credits, which are limited. I recommend everyone try it quickly.

4. Practical Guide to Cryptohunt API

During Vibe Coding, I chose the cost-effective Cryptohunt database. I didn’t master coding thoroughly but got familiar with the API, so here’s a quick guide.

Official link: https://pro.cryptohunt.ai/

1. Precise Feeding

Many beginners directly dump thousands of words of API documentation to AI, which causes confusion.

• Wrong example: Here’s the API documentation, you figure it out.
• Correct example: I want to develop a token monitoring feature. Please read the section about POST /tweet/mention_tweets in Cryptohunt API. Tell me how to get data from the last hour?

Prompt example: “Call the /tweet/mention_tweets API, search for tweets with ticker SOL. Limit to the last 24 hours. Help me filter out the top 5 with the highest interactions (likes + retweets) and see how people feel about SOL.”

2. Debugging神器

In Vibe Coding, the most frustrating situation is: I click a button, but nothing happens on the screen. At this point, beginners often doubt themselves: Did AI write wrong code? Is the network slow? Or is the API down?

You can open API usage to refresh:

• If there’s a record, and points are deducted: It means the API call succeeded! Data has been successfully sent back to your computer. This is a front-end rendering issue. You can tell AI: “Data has been received, but nothing displays on the screen. Please check the rendering code.”
• If there’s no record: It means the request was not sent at all! Then it’s a logic/network problem. Tell AI: “Clicking the button has no response, request not sent. Please check the click event.”

Just give the conclusion directly to AI; it will fix faster. During Vibe Coding, the real effort is no longer writing code but diagnosing where the problem is.

Once we delegate syntax and implementation details to AI, the focus of development returns to controlling information.

Therefore, everyone says that the end of Vibe Coding is not code, but data.

5. Final Words: Two Core Skills to Master Vibe Coding

After experimenting from Google AI Studio’s “trial,” through Antigravity’s “torture,” to Lovable’s “explosion,” I finally understand: tools are just swords in your hand; your vibe is the true Dao in your heart.

When trying Vibe Coding, remember these two core points—they are more important than any code:

1. Dare to Imagine: Imagination is your only “ceiling”

• In the era of Vibe Coding, the technical barriers have collapsed. Previously, we didn’t dare to make tools because we feared we couldn’t code it out. Now, practical experience proves that as long as the logic is coherent, AI can generate it.
• Don’t just be a “bricklayer,” aim to be an “architect”: boldly conceive products you once dared not imagine. A radar monitoring all online data? Or an automatic arbitrage robot? Your task is to turn that fuzzy “want” in your mind into a clear “blueprint.”

2. Be Good at Speaking: Prompts are your source code

Many beginners fail not because AI is stupid, but because they are too “polite” or vague.

Vague requirements = vague results

• Reject vagueness: Don’t tell AI “Help me make a good webpage.” (AI: What’s good?)
• Embrace precision: Be specific like when I made dashboards: “I want a dark mode dashboard, with these four data cards on top, call this API in the middle, and use AI to automatically summarize on the right…”

Remember, you are now the client boss. Your authority comes from “clear requirements” and “patient feedback.” If AI makes a mistake, don’t doubt yourself—just tell it: “No, that’s wrong. Redo. I want A, not B.”

If you’ve read this far, it means you probably already have the ability to create tools with Vibe Coding.

Don’t hesitate—turn your vibe into coding.