Tricks to start coding with AI for beginners

First Published: .

Last Updated: 2026-05-22 07:50 UTC+1

Some tricks for programmers that don’t know how to use AI tools, advice, pointing errors, best solutions…

General knowledge with AI and Code

The first thing to know is that the AI makes errors. They can make silly errors, like a missing + to concatenate strings in a varnish .vcl file or a random character like | between use imports in PHP or bigger bugs, difficult to spot bugs or security bugs. My advice is to not blindly trust the code generated by IA and to review it as much as possible.

An error generated by Claude inside trans
It displays wrong line numbers often. You can see in line 154 there is not what Claude points
It says "synClassroom" instead of "syncClassroom"

What the AI does very well, and it’s a great help, is to Review your Code. This is really helpful. Ask it to Review your Code before committing, or ask to review a colleague’s branch.

I git pull a branch and I use a prompt like this for Claude or Junie: “Do a code review for the new code in this branch respect main. Do not review files not committed“.

This way I don’t ask for a full Review of all the Code in the project, which is costly in tokens. I just ask for the new changes in this branch respect main. So normally, few tens of files. Quick and cheap.

AI it’s also very useful to determine the cause of an error. Just paste the error message, the error from the logs, or an screenshot and it’ll probably figure out what is wrong.

To be successful creating an application you have to provide detailed requirements. The most detailed, the better. The specs.

AI are very good assembling SQL queries. You can ask the AI to do SQL Queries, ALTER’s…

You can also ask for queries for specific CMS. Nowadays most AI’s are trained with the development documentation of CMS and Frameworks, so they know how to do stuff.

AI is terrible for Infrastructure. Do not give it the control of your Infrastructure, cause AI is not a clever human. It lacks common sense. If AI believes that the best way to update the infrastructure is to destroy your Production databases and redeploy them, it will do it. Even if you told it not do it. It may even delete the backups. The AI’s don’t obey always. Some times they forget the instructions that you give them.

For my new projects I always create a file named docs/promptia.md with the instructions to the AI, no matter if it’s Claude, Gemini, Junie, or another. I detail all the specs and requirements. This way I can reproduce the project again from the scratch.

There is a kind of standard AGENTS.md file, but Claude does not follow this.

Most AI use tokens as a measure of usage, to limit your usage. Is not exactly like this, but approximately one token equals one word in English. A word may take two tokens, and this may be less optimal in other languages. If you run out of tokens you cannot use the AI until you have more tokens again. Most of the subscriptions renew the tokens available monthly, weekly or daily. In some subscriptions you can pay for extra usage.

In most solutions you can choose to Plan before Executing, so to see what the AI agent will do. You can also use Ask to ask for questions about the code without modifying.

The context is the information that the agent has in memory about your project. The biggest the context is, the more tokens it will need, and the bigger probability that the AI will make errors or forget things.

The best engineer I know, working with huge code bases, limit the Claude context to a 200K window max, unless there is a specific analysis task for which they want to allow more. Otherwise the agent can spend a lot of time compacting. (Claude blog post: 1M context is now generally available for Opus 4.6 and Sonnet 4.6)

If you have to use AI, I recommend paying a subscription, as the degree of privacy of the information submitted is superior.

For security reasons I don’t recommend you installing plugins in the IDE’s, or skills in Claude (unless you read all of them and you know what you’re doing).

Claude Code

In my opinion Claude Code from Anthropic and Gemini from Google are the best IA tools, currently, to assist you in the generation and review of code.

Claude Code is an IDE available for Windows and Mac, for Linux as today you have to use the terminal application (which is quite powerful nonetheless).

If you use Linux, you can use Claude from PHPStorm, in the integrated terminal. You’ve to install the plugin. Follow the instructions from Claude’s web site. The integration with PHPStorm is very good.

From Claude in Linux (Terminal) you can also paste images. Just copy them to the clipboard and paste them. You can also drag and drop to the Claude Terminal.

See how I pasted an image, and Claude references it as [Image #2] as it is the second one that I pasted

When I create new application, I always create a file named docs/prompt.md where I detail all the requirements for the new application. Then I ask Clause to create the application based on this file.

When there is a Database involved, I always ask Claude to document the Schema in a document inside the /docs folder named docs/scheme.md.

You can type /model to select the model you want to use (some are more expensive than others, normally depending on how good they are).

You can navigate through the chat history of your requests by pressing the up or down keys in your keyboard (like in MySQL or Bash for example)

You can select if you want to work with the branch master, or another, or you want to work with Worktrees, which is very practical.

If you hit slash (/) and then hit the up arrow a few time you can see the chat.

/clear to start a new session and free the context. You can later come back to your old session.

/model to change the AI model using. Some are more clever, some consume more tokens. You can also change the effort. More effort provides better results, but also higher token consumption.

Videos:

Perplexity

I use it for general questions about Internet, as it presents all the sources and I like that.

I use also for questions about code, as a chat.

I pay 20€ per month approx.

ChatGPT

The most known. It hallucinates a lot. I don’t use it.

JetBrains Junie

JetBrains has their own AI agent that integrates in their IDE PHPStorm.

My company pays a $20 USD/month subscription.

Instead of tokens they use credits and this subscription provides 20.00 credits per month. In my opinion they get exhausted too fast.

Using Junie I saw it was getting confused with the dates. Not knowing when it’s today, and thinking that software already released (like Ubuntu 26.04) was not released yet (even if we were in May).

Videos:

Google Antigravity 2

The agentic tool. It is very similar to other vibe coding IA tools.

It is oriented to talking to the agent.

I preferred when we had the IDE and the agentic part together in Google Antigravity.

You can still still the Antigravity IDE as a separate tool.

Videos:

Google Antigravity

Antigravity was an IDE with support for Gemini IA agents integrated, and others.

As of 2026-05-20 Google presented Antigravity 2, which separates the IDE and the agentic tool.

It offers a great amount of usage of different models for free, however lately I see that their server are saturated often, and it’s impossible to use it for free.

As an IDE has the basic functions, integration with github, and can be a good alternative to PHPStorm if you don’t want to pay for the IDE.

It supports many languages, not just PHP like Python, Java… You can also create binaries.

Videos:

Android Studio

Android Studio is for creating Android APPs. It’s free and bring incorporated Google Gemini IA for free with some limits obviously.

I created several applications for free with Android Studio and Gemini.

Videos:

Github Copilot

It makes sense if your organization works with github, has the repos in there, and creates the issues in there, with good descriptions.

Then you can ask copilot to work on several issues, as if it was a member of your team. It will work in parallel. And you’ll receive an email when the task is completed, with a PR (Pull Request).

As I was not using Github at work, the usefulness was a bit limited for me.

Codex

The coding agent from OpenAI (ChatGPT).

Only available for MacOs and Windows. Not for Linux.

Views: 51 views

Rules for writing a Comment


  1. Comments are moderated
  2. I don't publish Spam
  3. Comments with a fake email are not published
  4. Disrespectful comments are not published, even if they have a valid point
  5. Please try to read all the article before asking, as in many cases questions are already responded