# Create your own virtual proto-agent: a guide to developing GPTs for any task

Author: Fernando Nieto Lobato
Original publication: 2023-11-29
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-9-personalice-su-inteligencia
English URL: https://elcontemplador.github.io/estrategia-english/essays/009/
Status: Published translation

This is a translation of the original Spanish essay published on 29 November 2023. Its claims, examples and forecasts retain that historical context.

English publication: 2026-09-29

*Historical context: this guide preserves ChatGPT's interface, account requirements and examples as described on 29 November 2023. It is an archive of that workflow, not current product guidance. Screenshot dialogues below are historical text translated from Spanish.*

After two information-packed newsletters — [the one on AI agents in Bill Gates's vision](https://open.substack.com/pub/estrategiabyaleph/p/estrategia-7-agentes-de-ia-redefiniendo?r=2tyybn&utm_campaign=post&utm_medium=web) and [the “Game of Thrones” at OpenAI following Sam Altman's dismissal and subsequent return](https://open.substack.com/pub/estrategiabyaleph/p/estrategia-8-juego-de-tronos-en-openai?r=2tyybn&utm_campaign=post&utm_medium=web) — **this week we want to bring you some straightforward, practical information so that you can create your own GPTs and customise them to perform tasks relevant to your work**.

Having created several GPT proto-agents internally for various tasks at [Institución Educativa ALEPH](https://institucioneducativaaleph.com/) and for this newsletter, I believe that these early assistants can substantially improve personal productivity and work quality when they are well defined and tailored to each activity's needs. They do not work equally well in every case, and it can often be difficult to find the precise instructions needed to achieve what you want. Even so, they represent another step beyond the prompt engineering we explore each week in our prompts for GPT-4 section.

**After some general instructions, we will work through a very simple example of creating a GPT through ChatGPT, step by step**. A paid account is required, as GPTs use GPT-4. We will also point towards a more advanced method that can produce systems with a greater degree of automation.

First, in case you are not yet familiar with what a “GPT” means in this context — although we referred to them directly in the newsletter on agents, since they are still very limited proto-agents — here is a definition: OpenAI's GPTs are customised versions of ChatGPT that users can adapt for specific tasks or topics by combining instructions, knowledge and capabilities.

## Instructions for creating a GPT

These instructions follow the tool's guided mode, which is the simplest approach.

1. **Access your OpenAI account**: sign in to your OpenAI Plus or Enterprise account at chat.openai.com. If you do not yet have an account, or use the free version, you will need to subscribe to a paid plan.
2. **Click “Explore”**: once you are signed in, click the “Explore” button in the menu on the left.
3. **Create a GPT**: in the GPT menu, select “Create a GPT”.
4. **Describe the chatbot you want to create**: use the GPT Builder to describe the kind of chatbot you want to build.
5. **Confirm your choice and choose a name**: the builder will present a summary and suggest a name for your chatbot. You can change the description or name.
6. **Choose a profile image**: while the builder prepares your GPT, select an appropriate profile image. You can use an automatically generated option or upload your own.
7. **Adjust your chatbot**: the builder will ask questions to customise it, such as how detailed its answers should be or what communication style you want it to use.
8. **Test your GPT**: try out your new chatbot to make sure it works as you want.
9. **Publish the chatbot**: if you are satisfied with the results, save it and decide whether to keep it private, share it with a selected group — as we will do here — or make it public.

**There are also more advanced settings** that we can define, including directly in another tab, “Configure”. For example, we can choose whether to give the GPT **the ability to browse the Internet, use the code interpreter or access documents we upload containing our information**. If you would like more detail on this direct method, [it is explained towards the end of this guide written by OpenAI's own team](https://help.openai.com/en/articles/8554397-creating-a-gpt).

## A very simple example

**We will create a GPT to write tweets — or posts on X, to use the new terminology — from any content supplied to it, such as a political speech or an election manifesto.**

Following the instructions above, we will create a GPT and **configure it through conversation: the simplest approach, although neither the fastest nor the most powerful**. We will give it the following instructions:

[![The November 2023 GPT Builder, with Create and Configure tabs, a preview pane and a Spanish prompt asking for an assistant that turns long content into five tweets.](https://elcontemplador.github.io/estrategia-english/assets/images/ddc3a4dd-3a84-433f-aa74-9bcadaba3b94_1600x807.png)](https://substackcdn.com/image/fetch/$s_!bmxN!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fddc3a4dd-3a84-433f-aa74-9bcadaba3b94_1600x807.png)

*Historical screenshot. The builder's welcome message is already in English:*

> Hi! I'll help you build a new GPT. You can say something like, “make a creative who helps generate visuals for new products” or “make a software engineer who helps format my code.”
>
> What would you like to make?

*Visible user prompt, translated from Spanish:*

> Good morning. First, I would like you to address me in Spanish while we create this GPT. I want to create an assistant that turns any long piece of content I provide into concise, engaging tweets.
>
> To do this, I want you to take a brief breath and, working step by step, write five tweets from each piece of content I provide, acting as an expert community manager.

We supply the initial instructions and it begins updating its “programming”. It then replies and **suggests a name for this proto-agent**. “Tweet Master” is clearly not the best name in the world, but it will do for this very simple example, which we are not going to configure exhaustively, so I tell it to go ahead. **It then creates a profile image for the agent**, using DALL·E 3 internally:

[![The GPT Builder proposes the name Tweet Master, generates a colourful profile image and displays an assistant for turning content into five engaging tweets.](https://elcontemplador.github.io/estrategia-english/assets/images/64456d28-f423-484c-b58e-6c455319ca27_1600x1098.png)](https://substackcdn.com/image/fetch/$s_!SfI9!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F64456d28-f423-484c-b58e-6c455319ca27_1600x1098.png)

*Historical screenshot dialogue, translated from Spanish:*

> **GPT Builder:** Perfect, I will continue in Spanish. I have configured the GPT to act as an expert community manager, transforming long pieces of content into five concise, engaging tweets.
>
> For the name, what do you think of **Tweet Master**? Or do you have another suggestion?
>
> **User:** Perfect.
>
> **GPT Builder:** Here is the profile image for **Tweet Master**. What do you think? Is there anything you would like to change or improve?
>
> **User:** That works for me.

*Preview text, translated: “Assistant for transforming content into five engaging tweets.” The conversation starters read “Turn this article into tweets”, “Turn this report into interesting tweets…” (truncated in the screenshot), “Create tweets from this text” and “Make five tweets from this content”.*

Finally, in this very simple configuration method, **it asks several questions to personalise the GPT further**:

[![The GPT Builder asks about content, tone and priorities; the user specifies political material and a warm, formal, respectful tone, and the builder confirms the configuration.](https://elcontemplador.github.io/estrategia-english/assets/images/f2f5ca49-9037-4017-8589-574b5612b3ad_1600x1441.png)](https://substackcdn.com/image/fetch/$s_!a5oa!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff2f5ca49-9037-4017-8589-574b5612b3ad_1600x1441.png)

*Historical screenshot dialogue, translated from Spanish:*

> **GPT Builder:** Great! Now, to refine **Tweet Master's** behaviour, I need to understand a few more details:
>
> - What kind of content do you expect to turn into tweets? For example, news articles, blogs or technical reports.
> - Is there a particular style or tone you would prefer for the tweets: formal, informal or humorous?
> - Would you prefer **Tweet Master** to focus on particular aspects of the content, such as key points, interesting quotations or statistics?
>
> These questions will help me ensure that **Tweet Master** responds exactly as you need.
>
> **User:** In this case, I would like it to specialise in generating tweets from “political” content: speeches, election manifestos, talking points, etc. I want a warm but formal and respectful tone that can also connect with people. I want you to use your knowledge to choose the aspects of the content you consider most relevant, so as to represent the material I provide accurately and position it well in public opinion.
>
> **GPT Builder:** Perfect, I have updated **Tweet Master** to specialise in political content, using a warm but formal and respectful tone. It will focus on highlighting the most relevant and strategic aspects of the content to connect effectively with the audience.
>
> Now that **Tweet Master** is configured, would you like to try it in the playground on the right? You can send it political content and see how it turns it into engaging tweets. If you need any adjustments or refinements, I am here to help. What do you think?

*The preview repeats the description and conversation starters from the preceding screenshot.*

**With these simple steps, our GPT proto-agent is configured. All that remains is to test whether it works as we want.** For a professional project, we would very probably use the “Configure” tab to make several more specific choices, but this is enough for a general example. **We click Save and make it available to anyone with the link.** In this case — although, I repeat, this is not a particularly refined agent, as we simply wanted something easy to configure — **I have run an initial test using Pedro Sánchez's full investiture speech**. These are the results it gave me from such a long piece of content:

*Archive note: the source introduces test results at this point, but the preserved article contains no text or screenshot of those results.*

**You can now try it**, if you are a paying GPT-4 user, **at [Tweet Master](https://chat.openai.com/g/g-jeR5u70sz-tweet-master)**.

I hope this simple example shows that creating a proto-assistant through ChatGPT is really easy. **They can also be created through OpenAI's website, but that route requires using the API, with its associated costs.** Alongside this very simple proto-agent, **over the past few days we have configured a somewhat more advanced assistant that we also want to make openly available — although it is still in beta — for all readers of this newsletter to try**.

## A speech generator that asks for parameters to create personalised speeches

[![The Orador Inspirado GPT's start screen, describing a tool for first drafts of political speeches and showing its activation button.](https://elcontemplador.github.io/estrategia-english/assets/images/80ad60ae-8312-44cc-80e1-09603882e9ac_1359x714.png)](https://substackcdn.com/image/fetch/$s_!gOsW!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F80ad60ae-8312-44cc-80e1-09603882e9ac_1359x714.png)

*Historical screenshot text, translated: “Orador Inspirado” (“Inspired Speaker”). “Specialist in creating first drafts of political speeches. Press ‘Activate Orador Inspirado’ to begin.” By Fernando Nieto. Button: “Activate Orador Inspirado”.*

**You can try it at [Orador Inspirado](https://chat.openai.com/g/g-3TQXKuTeE-orador-inspirado).**

**This one has more developed programming**, [following the method shown in this video](https://youtu.be/z8lMmODJ830?si=U0NzvN3uslpRJiyF), similar to other templates for generating agents circulating on Reddit. It allows us to create proto-agents that require activation and ask for a set of parameters by default. It can also provide some quite interesting instructions for developing your own projects, although we have made substantial changes to the initial draft to personalise and refine it.

## GPTs with documents and further examples

We have not specifically explored them in this article, but **GPTs that use documents included in their configuration are also particularly interesting. Here is one created by the journalist Kiko Llaneras about his book *Piensa claro* — “Think clearly”**:

[![Kiko Llaneras announces a GPT based on his book Piensa claro, with a preview offering explanations and examples of its eight rules for understanding the world through data.](https://elcontemplador.github.io/estrategia-english/assets/images/d1749a59-e672-4347-8b81-ddee1a445072_757x783.png)](https://substackcdn.com/image/fetch/$s_!7oN8!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd1749a59-e672-4347-8b81-ddee1a445072_757x783.png)

*Historical screenshot text, translated from Spanish. Kiko Llaneras, 12 November:*

> I have created a mini ChatGPT that is an expert on my book *Piensa claro*! And I am amazed. 😮
>
> How? It has read the book, and now you can ask it for passages, examples or explanations of any of its ideas. It works. Look:

*GPT description, translated: “I am an expert on the book ‘Piensa claro’ and its ‘Eight rules for deciphering the world through data’. What would you like to know?” The visible conversation starters ask why to read the book, how to apply one of its rules, what the eight rules are, and for an explanation of an interesting concept; three of the button labels are truncated in the screenshot.*

**[Try Piensa Claro GPT](https://chat.openai.com/g/g-NVDEG0xuh-piensa-claro-gpt)**

**Finally, if you would like to explore further, many GPTs are being made fully public**, although relatively few are really valuable at present. **To find those likely to be better, we recommend this GPT directory: [Discover the Best GPTs](https://supertools.therundown.ai/gpts), from The Rundown.**

Fernando Nieto Lobato

*Director of Digital Innovation, Institución Educativa ALEPH*