Getting Started with AI Programming for Beginners
Personal opinions, for beginners’ reference only. Please forgive any errors or omissions.
In most people’s minds, AI may only be capable of generating text, images, and videos. In fact, its capabilities go far beyond that: recognizing everything, intelligent editing, program scheduling, information retrieval, data analysis, logical reasoning… and much more is possible.
Of course, what fascinates programmers is how AI can serve as an assistant for writing programs, saving time and effort. Even someone like me, who never attended middle school, can now use AI to write projects.
I saw a project by an expert: a website directory commonly used by programmers. Its advantages are that it reduces the need to maintain bookmarks and makes it easier to find the login pages of various sites. The address is https://nav.eooce.com/ Lao Wang’s Navigation Site.
So I created another navigation project by imitation and deployed it to GitHub. The address is
白狐 White Fox
It is a navigation website I developed specifically for programmers and related enthusiasts. Copy and paste the URL into your browser and click Go to open it; do not search for it on Baidu. Enter the domain name and press Enter. If you only enter the domain name, the browser will automatically recognize it and configure the HTTPS protocol and port—https://—so you do not need to add them manually.
However, because my site has limited bandwidth, I still recommend that programmers deploy and use it themselves. This reduces the bandwidth usage on my end. Although GitHub provides 100 GB of traffic per month, the site still frequently goes offline for various reasons and is not very reliable. My repository address is https://github.com/757926525/daohang/settings/pages
. It supports deployment on various mainstream platforms. If it opens a 404 page, use this link: https://github.com/luobikang1?tab=repositories to search.
This provides an approach for a project. Everyone can create different navigation sites according to their needs—for example, asking AI to add the sites they want while programming.
I looked through the free options on various mainstream platforms and tried several of them. Personally, I think replit.com is quite good. Although many AI tools can now connect directly to repositories to save files, something that used to be free, some now charge for it. This service is the same, but it has a workaround: tell it your repository address and token, and it will push and save the files.
This is indeed troublesome, but if free things were not troublesome, who would still pay for them? Of course, if you do not need to save things to a repository, it also provides a deployment project archive. The prerequisite is that you explain this to the AI.
AI programming consumes a great deal of resources, so free plans have fairly strict limits. If you do not know how to use them, you can quickly use up your daily credits in just a few attempts.
For an ordinary project, if you are technically proficient and use relatively little dialogue, it may be possible to finish it in three to five days.
The simpler the prompts, the better, but they must have a basic framework: which functions are essential, which things must not go wrong, and which parts need polishing. You must tell the AI all of this; otherwise, even if it has the materials, it will not know what to do.
Example:
Generate a film and television project with basically the same functions as Moon TV, but make the video resolution adjustable down to 360p. Support pull-and-deploy and upload-and-deploy. Support deployment on Cloudflare and various other platforms. Support deployment to Cloudflare Pages, with the ability to add an API in the settings.
(This describes the main functions.)
Optimize the project and fix the issue that prevents access after deployment to Cloudflare. Change the panel name to White Fox. Require a password to open it. Use D1 to synchronize data. Add deployment instructions and notes to GitHub, and highlight key variables with boxes. Check for errors and ensure that deployment succeeds through all available methods.
(This is error correction: you must tell the AI what is missing, what has been added, and what is wrong.)
Finally, provide Pages deployment files that can be uploaded directly to Cloudflare. Provide project files for upload-based deployment, save the project files to GitHub, and save them to GitHub for downloading.
(This makes the project more complete and able to adapt to various methods.)
Of course, AI is not a person. Even the best AI may achieve nine out of ten points, while that last point is always unsatisfactory. Therefore, once the basics are complete, you must deploy the project to various platforms for testing and revise it whenever you are dissatisfied.
Impressions of each platform
- Cursor: unclear.
- Replit: supports deployment within the platform and is moving toward comprehensive development. It has problems connecting to GitHub.
- Qwen Code: unknown.
- OpenCode: a company with global potential, but it has a high barrier to entry for nonprofessionals.
- Gemini: backed by Google; an old hand taking advantage of newcomers. It can be used directly with a Google account through authorized login.
- Claude: a major promoter of machine intelligence. Because too many people use it, it restricts users in countries with large populations to reduce costs. Its system automatically detects anomalies, and accounts may be banned without explanation. It has problems connecting to GitHub.
- http://jules.google.com/ An AI service under Google.
The mobile interfaces of Claude, Jules, and Replit are broadly similar. Registration generally requires a Gmail address; other details are unknown.
Devin has a polished interface and relatively few errors. However, its $20 trial allowance can be used up after only a few rounds of questions and answers.
《Reference material》
Claude is a cutting-edge generative AI assistant developed by the AI company Anthropic, powered by a powerful large language model (LLM). It can not only answer questions, write articles, and generate ideas like traditional chatbots, but also offers exceptional long-text reading, programming assistance, and data analysis capabilities. Claude’s three core working modes As AI technology has evolved, Claude has long surpassed ordinary “chat” functions and developed three primary usage modes: Think tank (Chat): The classic question-and-answer mode. Suitable for brainstorming, quickly acquiring knowledge, summarizing documents, and conducting in-depth business logic analysis. Automated collaborator (Cowork): Combines automated control capabilities and can directly connect to cloud drives such as Google Drive or project management tools such as Notion. It can handle tedious classification, aggregation, and organization tasks for you. Development mode (Code): Supports tools such as Claude Code. It allows users who do not know how to program, as well as developers, to run and modify real files directly in the environment and even complete multi-task development workflows fully automatically.
Claude’s four major strengths
Extremely powerful long-text processing: Supports a very large context window (Context Window), allowing you to submit dozens of papers, an entire lengthy contract, or hundreds of thousands of lines of code at once. It can accurately identify key points and produce cross-file comparison tables. Natural and fluent Chinese: It can understand and generate both Traditional and Simplified Chinese very well, while accurately imitating a specified tone and writing style. Ready-to-use output: It can directly create and output files in formats such as Excel, PDF, and Word, or produce interactive charts and small tools that run independently. Custom Skills: Allows users to package specific workflows into dedicated skills (Skills) for one-click automation. Yes, Claude can connect to GitHub. The official service provides a dedicated GitHub integration and Claude Code GitHub Actions, allowing you to seamlessly integrate AI programming capabilities into your development workflow.
There are two main ways to connect and use it: 1. Direct integration in the web or desktop version (GitHub Connector) You can connect a GitHub repository directly to Claude’s web or desktop version, allowing Claude to gain an in-depth understanding of your codebase: Multi-project knowledge base (Project Knowledge): Click the “+” in Project Knowledge, select GitHub, and search for or paste your repository link. You can connect public or private repositories (authorization required). Directly reference in a conversation (Chat): Click the “+” in the lower-left corner of the chat interface, select “Add from GitHub,” and directly load specified files and folders.
Usage details
Advantages: Claude can read project content as context and supports one-click synchronization to obtain the latest code changes. Claude Free’s usage allowance is based on a five-hour period. The exact number of messages varies dynamically according to server congestion and conversation length. Message limit: Ordinary conversations typically allow around 10–15 messages every five hours. If your questions are complex or the context is very long (supporting up to 1 million tokens), the allowance will be used up more quickly. Reset time: Once the allowance is exhausted, you must wait five hours for it to reset, or upgrade directly to Claude Pro for $20 per month, which provides approximately five times the message allowance. How are messages “actually counted”?
This is the source of confusion for most people. A “message” is not a single chat bubble—it is measured by the number of tokens consumed:
A short question (“Fix this spelling mistake”) = ~200 tokens → barely affects the limit
A medium request with code context = ~5,000 tokens → moderate consumption
Claude Code reading 10 files and generating modifications = ~50,000+ tokens → heavy consumption
This means that 45 short Claude.ai chats ≠ 45 autonomous Claude Code operations. Heavy Claude Code usage consumes the allowance 5–10 times faster than ordinary chatting. logo
Among the many free CLI tools, Qwen Code chose to start charging
When it first came out last year, it gave away 2,000 calls per day with no token limit—you could use it freely. At the time, everyone in the group was praising it, saying that getting this thing for free was incredibly worthwhile. Some even said, “Claude Code can be thrown away.”
To be fair, the free allowance was indeed astonishing at the time. Two thousand calls a day was more than enough for normal development.
Then things began to change.
First, the Lite plan stopped accepting new purchases on March 20, leaving only the Pro plan at 200 yuan per month. Then the free allowance was cut from 2,000 to 1,000, and now it is down to just 100 calls. Honestly, 100 calls are only enough for a brief greeting; write a few lines of code seriously and they are gone.
I understand commercialization, of course. Free services cannot last forever. But the problem is—look at the competitors around it. OpenClaw also has restrictions, but at least its free allowance remains usable; Codex CLI is open source, so running it locally costs nothing; Cursor has a free tier, and Claude Code offers pay-as-you-go API pricing with clearly stated rates. Everyone is trying to lower the barrier to entry, while Qwen Code is doing the opposite, tightening restrictions more and more.
，分享点赞