A guide to programming with AI when you can’t program

How to program with AI when you can’t program

AI has changed one more thing. Suddenly, even non-programmers can program. And that’s what we’re going to look at today. So this article is not for programmers to increase their productivity, but for non-programmers who need to create a program from time to time.

Programming has a major disadvantage over many other creative practices: you have to know something about it. If you open a development environment, you have to know how to compose commands for yourself, and until you get it right, nothing works. If you want to write or draw, you might not write or draw anything particularly pretty, but somehow it will work. With programming, it doesn’t - you either know how to do it, or you don’t do shit.

But AI changes that. If you subscribe to ChatGPT or Claude, you can let AI create even very large programs for you. Of course, it’s better to start with smaller scripts and get a bit familiar with the whole thing.

Let’s take a look at how to do it.

First, you need to be clear about what you want to create. Definitely don’t assume that the AI will generate Facebook-scale sites for you in a single prompt. First of all. Secondly, it’s good to know a bit about web development and operations, obviously the more the better, but any basics are useful. Thirdly, think of AI as a more experienced colleague that you can ask for, but you have to explain it well.

The process of building an AI program is an iterative process. You have an input - that is, a baseline and the conditions from which it starts. Then you have the output - what you want to get and what you want it to look like. And in between is the magic that laymen tend not to care about, which we now hope AI will solve.

Let’s take a task. For example, I use a system for logging projects where I write down what I did and follow it with a hash of the name of the project or work. For example, “Ordered new server #marigold#. And I want to have a web service where firstly I note this down, secondly I can list all these notes. It’s a relatively simple application for our AI programming example.

First, let’s be clear about what I want:

the data is entered via a web form in arbitrary short text format and can include a hashtag, a word starting with the # this data can be output by date and also limited to a specific hashtag as this is a web service accessible from the internet, it will be appropriate to make the data available only to the logged-in user

And that’s actually the whole brief of our project. Deliberately ripped off, which an experienced designer would rightly accuse me of, but we’ll see why later.

Let’s assume you’re paying for ChatGPT, but the procedure will be similar for other tools like Claude. In ChatGPT, however, you have the advantage of being able to use GPT. Find a tool called Code Copilot in the Explore GPT section, it’s a great customized tool for programming support. Let’s insert roughly this prompt:\

Hi, I need to create a web application. The application should allow user registration and login. The logged in user can enter text in the text input field, which will be saved. It is saved including the date and time of entry. Another page allows to list all the entered texts in a selectable time range and sort them by hashtag. The next page generates a cloud of hashtags including frequency of use. I expect the service to be underused. What programming language and approach do you consider most appropriate for this assignment? I would like to produce readable code and you can help me choose a hosting where I can easily deploy this code, everything will be readable and understandable for a beginner. I would like the hosting to be free or paid depending on the frequency of use. First, give me a summary of your recommendations for programming language and for hosting, describe the pros and cons.

Code Copilot (CC) will offer to implement the solution in Python, using the Flask framework for the backend and Jinja2 for the frontend. This is where it comes in handy if you know roughly what’s going on, we won’t explain.

CC will also recommend platforms for hosting Python, he suggests using Heroku, alternatively Vercel, Netlify or Render, all very valid choices. But how come he didn’t recommend, say, Amazon AWS or MS Azure when there are ads for them everywhere? Let’s ask:

Why didn't you recommend AWS, Azure or Google Cloud?

CC replies that these are very complex platforms that will be difficult for a beginner to set up. Heroku, Vercel, and similar platforms are focused on simplicity and user-friendliness, which is good for lower-end projects and for users who don’t want to deal with complex configuration. Which I can attest to, setting up AWS/Azure is tedious for a beginner. By the way, Claude provided this disclaimer on that very first prompt. We provided the fact that we are beginners to AI in the initial prompt and immediately defined that ease of deployment would be important to us.

We can still ask how the hosting situation is in the Czech Republic (I get recommendations for Active 24, Wedos and Stable, all relevant) - but AI correctly reminds me that in this case I should be looking at a virtual private server, which is more expensive and already requires administration experience.

So, let’s put the first indulgence to bed and figure out the next workflow to get everything set up:

What would you recommend as a workflow for developing this application of mine? So I'm going to use python and Heroku, and my computer is running macOS.

Both CC and Claude give us pretty much the same advice (with varying degrees of detail) on how to proceed.

First, we should set up a local development environment, i.e. install Python, Flask and some source code editor - we are advised directly what and how. That’s the reason we put in the prompts what operating system we will be working from. AI now directly offers us command-line commands or software suitable for that operating system.

If we don’t understand something, we can ask the AI directly. For example:

How do I install Git for version control on macOS?

And I’ll get the details I need.

After some time following the AI’s instructions, we should be pretty well equipped. We have a local development environment so we can develop and test the application right on our machine. We have a versioning system in place, so we have good control over how we create the code and can deploy it “to production” at the appropriate time - or send it to Heroku and make it visible to the world there.

Now we’re clear on what we want and how we’re going to proceed. At this point, the “context window” is important, i.e. the fact that modern LLM models keep long texts of the previous conversation in memory. Thanks to this, we can just give a simple prompt to the AI to generate the application as we have designed it:

Create the source code for my app and describe in beginner's detail what to set up where and where to save it so that the code works when deployed.

The AI will now describe the file structure, i.e. where to store what, and then output the contents of those files to you. Now comes the tedious click-and-click moment where you have to use Copy&Paste to transfer the code (preferably via that installed development environment) to your computer and stack it where you have the project directory in GIT. And now comes the moment when you can test the application on your local machine. Only, how? Let’s ask the AI:

How do I test the application on my local machine running macOS?

You’ll get a description of how to test the app in your environment and on your computer. It is very easy for the first run of the application to fail for some reason. In that case, an error message will appear on the console. Copy the message and prompt:

I got this error message: and copy&paste the error text or log extract

The AI is quite flexible here, you can copy a very long error log and it will figure it out. It will explain where the error occurred and suggest an adjustment you need to make.

Once you’ve debugged the app, deploy it to Heroku and test it again, there can be errors here too. Again, you consult the error log with the AI.

You can probably see where the main problem is. The constant copying of source code from the AI response to the development environment. Fortunately, most AIs will format the code nicely and offer you a button to easily transfer the text to a clipboard, but it’s still a hassle. It would be better if the AI could directly link this output to Github or your GIT, but you can’t do that yet, you have to save each file one by one.

Tips:

Sometimes the AI will recommend you to make a specific fix in the code, for example just rewrite a piece of functionality. If you can find it in the code, do it, if not, tell it to give you the whole corrected code. Other times, you might be tempted to fix something in the code yourself, like text - in that case, tell the AI, because the next time it fixes the code, it wouldn’t know about your changes.

If you don’t understand something, have the AI explain it to you, ask.

As you can see, the approach is not entirely suitable for large-scale application development.