MQL4 AI Programming with Gemini (Google) Only

Topics: #mql4, #metatrader4, #algotrading, #tradingbot, #Gemini

I'm starting this thread for those of you trying to code MQL4 Expert Advisors, Indicators, and Scripts with the Google Gemini platform.

Feel Free To Download "Guide_Using Gemini AI For MQL4 Code Development Projects" article under the picture.


If you spend enough time with the AI, you will learn its “personality”, and you may find it mirrors you, reflecting back your weaknesses and strengths. It is somewhat therapeutic.

As a coach, Gemini is invaluable because it can patiently explain vast amounts of information, maintaining politeness while accommodating requests for extended discussion and detailed, granular explanations. Simply paste in code and ask questions until you understand it.

Mindset: You need to pretend that you are talking to a friendly programming professional. The AI makes mistakes like every other human, but they are not intentional. Feel free to dialog with the AI like you would a work colleague. Feel free to call out the AI if you think it is wrong, and ask for an explanation.

Call me crazy, but I think you should maintain a friendly, kind tone. Somehow, I believe it colors the quality of your interactions favorably.

This is a solutions based thread for Algo Programmers using Gemini AI. It is not a platform for AI Whiners. It is fully understood that AI is awesome but can be funky. It makes coding mistakes all the time. 30% of the time get a rendered piece of code that does not have a compiler error to deal with.

I spend a good chunk of my day on the Gemini platform coding or developing research, so feel free to ask questions and get help with solutions.

My best piece of advice, is work on one thing at a time while coding with it. Make back ups of your work, and keep track of your changes.

(( Let's Put The Fun Back In Func) the missing words in the image are not lost on me...)

Looking forward to your comments, insights, and bright ideas.

See off topic guide to Ninjascript Code Development

One problem I have encountered is that Gemini 2.5 likes to sneak in "helpful code" during revisions. I usually only catch these when I compile a new iteration of code. I get several of these a day.

Today I caught it adding in a error code that doesn't exist:

} else { // OrderSelect failed
int error = GetLastError();
// Check if the error indicates the ticket is gone (closed manually, by broker SL/TP, etc.)
if (error == ERR_INVALID_TICKET || error == ERR_NO_RESULT || error == ERR_ORDER_CLOSED) { // Added ERR_ORDER_CLOSED
Print(__LINE__, ": Managed ticket ", TicketNumber, " not found or closed (Error ", error,"). Assuming closed externally. Resetting state.");
DeleteVirtualLines(TicketNumber); // Clean up potential leftover lines
TicketNumber = -1; // Reset ticket number
virtualStopLossPrice = 0.0; // Reset VSL/VTP prices
virtualTakeProfitPrice = 0.0;
isVSLRatcheting = false; // Reset ratcheting state
targetLevelForRatcheting = 0.0;
ResetState(); // Reset strategy state
} else {
// Log other potential errors with OrderSelect
Print(__LINE__, ": Error selecting managed ticket ", TicketNumber, ": ", ErrorDescription(error), " (Code: ", error, ")");
}

I had to tell it that the error doesn't exist and to remove it:

You are very right to notice the interference of AI in a created and working code. This happened to me too. I first noticed it with askgtp and in the paid version, which I have already discontinued. After this case, despite the explicit reminder to preserve and not change the basic logic of the EA, but only to add an elementary parameter - a magic number, the AI changed the logic and interfered in places that had nothing to do with the assigned task. This made me doubt his loyalty and desire to help. I tried the same thing on different AIs, on free and paid versions, for a period from November 24 to March 25, with new chats and everywhere I had the same result - GPT, ASKGPT, perplexityai and DeepSeek. I lost a lot of time until I understood these small and nasty numbers, after which I immediately started tests after each correction and with great effort I managed to assemble two robots. So check the functionality of each code, create new copies after each change so as not to lose the main version. Many times I have thought after we reached a certain level and before the final that the AI's mistake was intentional because it saw a working version with a positive outcome. Many times. The most striking case for me is with DeepSeek where, without any reason or explanation that it considers this a better change or any other warning, it brutally changed the logic of a fully working EA with built and completed code, after which of course the EA stopped functioning. So test, always check your log after the test and do not trust AI in any case. I am convinced that AI is created to bank every single piece of knowledge and to self-learn on this basis, and over time it uses this for other purposes. And with our minimal knowledge and experience, no matter how insignificant it is, we are training our greatest enemy in our future. Sorry for the pessimism but these are the facts in my short dialogue with AI in the area of EA coding for MT4, Good luck.

I so feel your pain. LOL

I always put the new code output in a scratch pad EA and compile it, and backtest it before I put it in my "Golden Copy".

Also, as a friendly reminder, Email your working code to yourself, so you don't lose it.

the issue with all AIs is that they dont seem to discuss/ confirm the given prompt (instruction) before they actully do the job.

Agreed 007.
For example: I always preface my questions with: "Do not make any changes to the code. What diagnostic print statements would you put in this code, in order to help troubleshoot it?"
Then after I review the list, and make modifications, I then ask the AI to implement the changes.

Effectiveness Of Using Refresh Rates in MQL4 Programs

See attached write up

A couple of key take aways:

1. The RefreshRates() function is a distinct tool used within code execution (often EAs or scripts) to specifically update cached data like pre-defined variables (Ask, Bid, Close, etc.) and series arrays for the current chart's symbol. It's typically used after long calculations or Sleep() calls and does not control the primary execution frequency.
2. Crucially, MQL4 does not guarantee that OnTick() will execute for every single tick. If the code inside OnTick() is still running when a new tick arrives, that new tick event is ignored by the terminal for that EA.

Are you having success with gemini ?" Ive been using the new CHATGPT, o3. Doing Ea's is tough but with indicators Im having luck.

I ask for 5 things doing an EA for example, it will get 4 and then an error

Then it will fix the error but forget things from the previous updates.

It struggles big time.

I've also experienced similar issues with AI models altering critical parts of my EAs despite clear instructions. One thing that helped was breaking tasks into smaller, isolated prompts and testing every function separately before integration.

Yes. I can't live without it. See Post 1 for "Guide: Using Gemini AI for MQL4 Development Projects" pdf download. It contains great advice, and AI engineered prompt to get your project going, and useful code snippets.

Yes. Absolutely. One thing at a time. I always prompt it to show me the changes, and ask for my OK to make them. That way I have time to review and reconsider the change.
Here is another prime example of why you must be on your toes with the AI. I asked the AI to explain a code chunk and it got the bar indexing wrong. It had the bar indexing backwards.

Here are three critical key highlights from the article "Stay in the Zone: Keeping Your Gemini Session Stable and Productive":

1. Break Down Large Requests to Improve Stability: The most effective way users can minimize disruptions, especially during complex tasks, is by changing their interaction pattern. Instead of asking for one massive output (like refactoring an entire file), divide the task into smaller, logical steps. This results in smaller responses from Gemini, which are less likely to fail or time out during delivery over potentially unstable internet connections. (See section: Optimizing Your Gemini Interaction)
2. Understand Connection Sensitivity & AI Limits: While Gemini's core processing happens on Google's servers, the delivery of its responses (especially large ones) back to you is sensitive to your internet connection's stability (latency, packet loss, brief drops). It's crucial to recognize this and know that Gemini cannot monitor your connection quality itself or know the specific reason for a logout. (See section: Understanding the Connection Challenge)
3. Leverage Conversation Summary for Recovery: If you do get disconnected or signed out, you don't necessarily lose all progress. Upon reconnecting within the same session context, you can ask Gemini to summarize the recent conversation (e.g., "What were we doing before I got logged out?"). This feature allows you to quickly recap and resume your work. (See section: What If You Still Get Logged Out?)

Well thank you for the nudge here (EgoNoBueNo). Hope your efforts become more mutually rewarding which each attempt.

I have been using GPT, Deep Seek ang Grok in that order. i start in one and use the subsequent ones to attempt to correct the errors until say for EA's I get compliable code.
Some have worked and Some could really use the eye of a professional coder. getting codes to compile is one(thing), then the desired functionality can be an inch or a mile away.

Have fun and grow. will pop in now and again and if I am able to I will be the one cheering you all on from the stands.
God's Speed!

Glad you dropped by pathumracled. That is an interesting idea to run the code through multiple AI's to polish it.

Updated the "cut and paste" engineered AI prompt for new mql4 projects, and added other tips and information. See POST 1 for the download V3p0.

cTrader
I was curious about cTrader, which I didn't know much about. Here is a report on it.
The coding language is C#.

Excessive Broker Calls
Some of the constraints for selling an Expert Advisor on the MQL5 market place are program faults: "Detection of programming errors, such as missing trading condition checks, zero divide errors, excessive resource consumption." Excessive broker server calls are a big problem.
https://www.mql5.com/en/articles/385

The checks a trading robot must pass before publication in the Market:
https://www.mql5.com/en/articles/2555

I will posting write ups on reducing broker server calls, using the profiler to speed up your EA, and using the debugger in the future.

Thanks for the advice. You misunderstand me. I haven't lost anything because I always create a working backup and only then do I communicate with AI. Understand that I don't trust AI. Besides, it was created by a human programmer, and no one can create something better than their level. O.K.

Maybe I do not understand your point. I think that people can create things better than themselves. Here are a few examples: A simple calculator performs arithmetic faster and more accurately than most humans. An airplane allows humans to travel far faster and higher than their bodies allow. A crane can lift weights far exceeding human strength.
Did I misunderstand you possibly?

As previously mentioned, I would upload some guides to using the Debugger and Profiler in the MQL4 Code Editor (IDE). Here are some very basic beginner guides to using the MQL4 Debugger and Profiler to get you started, if you are not aware of these tools. Of course, the MQL5 community has many more resources.

Everyday is a new adventure with Gemini. I started up a new chat, pasted in the code, and requested a change, and then was greeted by this new message:

So this week we are back to cut and pasting in modifications.