In this article, I explain how I use prompt engineering with Codex, AI agents, master prompts, text files, and a private server to organize, develop, and improve my web projects.
For some time now, I have been using a prompt engineering strategy, or prompt engineering in French, that I find particularly interesting for moving forward with my web projects using artificial intelligence.
This way of working has developed gradually. At first, I started with a shared hosting server for various web projects, particularly around Wikidata. I have talked about it on my profile and on my blog. It required a lot of time and a lot of effort.
But I eventually reached a fairly simple limitation: the shared hosting server no longer allowed me to implement some of my projects the way I wanted to.
So I started reviewing my projects, keeping the ones that seemed the most reasonable to me, and returning to a private server that I had already had for eight years.
That is when my use of Codex, agents, master prompts, and text files really took on another dimension.
From a shared hosting server to a private server to use Codex
I had kept this private server for about eight years. I gradually optimized it to use Codex from the command line.
My goal was not necessarily to directly use services that already offer an organization built around agents. I preferred to start by creating my agents myself and organizing my projects in the way that suited me.
The logic is quite simple.
For example, I can ask ChatGPT:
- create an agent for this;
- create an agent for that;
- structure an agent for me with this feature;
- help me prepare a master prompt;
- summarize the information related to a project.
I also sometimes use voice dictation to prepare these requests.
Then, I can ask Codex to develop from this structure and gradually adapt the agents or master prompts according to the results obtained.
I am deliberately skipping the complex details related to web development here. What interests me most is the method.
On my server, this method relies heavily on text files. But the principle also seems usable through a web interface.
However, I would like to clarify one important point: I mainly work this way on my server. Therefore, I cannot confirm exactly how the same organization works on every other service that I have not tested.
Creating my agents and master prompts with ChatGPT
At first, I created my master prompts and agents with ChatGPT.
I then saved them as text files in different Codex directories related to my web projects.
This organization allowed me to start using Codex more and more intensively to develop several projects.
Today, some of these projects are still in the testing and refinement phase.
However, not all projects have followed the same path.
I have had to set some of them aside, particularly because of:
- a lack of disk space;
- a lack of power;
- server capacity;
- the time needed to continue developing them.
Conversely, some projects can be kept, improved, monetized, and gradually presented by me.
That is also what interests me about this organization: I can see more clearly which projects I really want to continue.
Listing agents by project
Once the agents have been created, one step that I find particularly useful is simply asking for a summary of what already exists.
For example, you can ask:
- list my agents for my different projects;
- list my agents by project;
- summarize my master prompts by project;
- summarize my agents by project;
- give me a summary of my master prompts and agents.
This step may seem very simple, but it allows me to find what I have already created more easily.
When several agents, several files, and several projects begin to accumulate, the summary itself becomes a working tool.
Instead of starting from scratch every time, I can reuse what already exists, see how the different elements are organized, and continue working from that base.
This is where, for me, prompt engineering no longer consists only of writing an isolated request.
It also consists of working on prompts that have already been created, revisiting them, summarizing them, and adapting them.
Using documentation to refine agents
For my web projects, I also have various documents.
Some were produced directly by me. Others may have been produced with the help of artificial intelligence.
In web projects, you will notably find README files, those “read me” files that contain information allowing you to understand how to use a system or a project.
These documents can also become useful for working on agents.
You can ask ChatGPT how to use a system based on this documentation, but you can also ask it to produce documents that can then be used to refine the agents and master prompts.
For me, this is also part of prompt engineering.
So I do not limit myself to a succession of questions and answers.
I can work from:
- my agents;
- my master prompts;
- my text files;
- my README files;
- documents that have already been produced;
- information gradually added to my projects.
All of these elements can then be reused and organized.
Harmonizing master prompts and agents
Once I have obtained a summary of the agents and master prompts for my different projects, I can begin another task: harmonization.
The idea is to look at what already exists and ask artificial intelligence to help me make everything more consistent with my objectives.
This harmonization can address different needs.
I can try to:
- spend fewer credits;
- be more efficient;
- be more productive;
- better organize my different projects;
- make better use of the information already available.
I can also add information or files, then revisit the master prompts and agents based on this new information.
On my server, part of this work can be automated using different tools. Once again, I am deliberately skipping the technical details because what matters here is mainly the overall logic.
I create, I save, I summarize, I add information, then I harmonize.
This sequence of steps is now an important part of the way I work with Codex.
Paying attention to Codex credit usage
Another factor to take into account is, naturally, the use of Codex credits.
Depending on your usage, you can see whether you are using the service too intensively compared with the credits available to you.
If that is the case, you can try to optimize your system so that it consumes fewer credits.
Obviously, not every use requires the same amount.
Text processing is generally relatively light in my use. However, if you start having entire books processed or produced, the amount required will naturally be greater.
This is why harmonizing your agents and master prompts can also have a very practical benefit.
You can ask yourself:
- do my agents still correspond to my projects?
- can some master prompts be grouped together?
- are my documents being used correctly?
- can I obtain the same result in a simpler way?
- is my organization still suited to the credits available to me?
The goal is not necessarily to reduce everything.
The goal is above all to adapt your organization to your actual usage.
Why my VPS has become interesting for my web projects
Today, this strategy allows me to use Codex very regularly and move forward with my business projects with greater confidence.
Moving from a shared hosting server to a private server required a lot of work.
For several days, I had to rebuild my environment and transfer my projects.
But I can now clearly see the difference between the two approaches.
To summarize my situation:
goodbye shared hosting server, hello private server.
My VPS is particularly interesting to me because it is based on a monthly plan and because I encounter fewer limitations than with my former shared hosting server.
What is also interesting is that today I am doing things by myself that I did not necessarily know how to implement in this way eight years ago.
I had nevertheless completed advanced studies in this field.
But I did not always have the courses, notes, or records of everything I had learned previously.
Today, artificial intelligence allows me to transcribe this knowledge much more easily and reuse it.
Things I learned a long time ago can therefore return to my current projects in a form that is much easier to use.
Can you use this strategy without being an IT professional?
I personally work on a server and with Codex from the command line, but that does not necessarily mean that you have to reproduce exactly the same environment.
Services such as Claude Cowork or OpenAI Work can, in my opinion, be interesting for people who are not IT professionals.
However, I have not tested all of these solutions, so I prefer to remain cautious about exactly how they work.
In any case, using agents and master prompts requires a minimum of:
- learning;
- developing your skills;
- understanding what you have already created;
- working on your own needs.
The technique I am sharing can nevertheless remain the same in principle.
You start by creating your agents.
You create your master prompts.
You keep the important information.
Then you ask for summaries by project.
You use your documents to complete these summaries.
Then you gradually harmonize your agents and master prompts according to your objectives.
My prompt engineering strategy in summary
Today, my prompt engineering strategy with Codex is therefore based on an organization that is fairly simple to explain, even though setting it up can require work.
I start from my web projects and gradually build around them:
- agents adapted to each project;
- master prompts;
- text files stored in my directories;
- README documents and other useful information;
- summaries of agents and master prompts;
- gradual harmonization of the whole system.
I can then continue improving this organization according to my projects, my server capacity, and my use of Codex credits.
In my opinion, prompt engineering therefore does not consist only of finding “the right prompt”.
It also means working over time with the prompts, agents, documents, and information that I have already produced.
This is the method I currently use to continue developing, testing, and refining my web projects.
It has also allowed me to better distinguish between the projects I needed to set aside and those I could continue developing, keep, and potentially monetize.
After several years with my private server and a period using a shared hosting server that had become too limited for some of my needs, I am ultimately rediscovering the full value of a VPS combined with Codex and artificial intelligence.
And if you also use agents, master prompts, or Codex, I mainly hope that you find the organization that matches your own projects, your usage, and your budget.
And, of course, that you do not spend too many credits if your budget is limited.
José DUTIAN