Basic tricks to know how to give orders to an artificial intelligence and thus generate better images and texts

ChatGPT (a conversational AI that provides results in text mode) has something in common with MidJourney or DALL-E 2 (image-generating AIs): the way we can use them to generate the content we are looking for is through instructions in natural language, something within the reach of any user. Of course, by ‘natural language’ we only mean that we don’t use programming languages…

…because, on the other hand, we must be careful how we speak to them if we want to be ‘answered properly’. First of all, we have to know how to condense the fundamentals of our query, because the AIs will ‘go ahead’ with the order we enter, without asking us for examples or clarifications. In addition, there are some instructions that cause specific reactions from the AIs, and we have to know ‘what keys to press’ (in a metaphorical sense) to get them.

Being an expert in this field has already become a profession in its own right: instructions engineer (or ‘prompt engineer’, if we like anglicisms). However, there are some basic tricks that we can learn to handle ourselves in this field.

For image-generating AIs

• keep it simple: use clear and concise language to describe to the AI ​​what you want it to do. Of course, ‘simple’ is not ‘short’: using less than 7 words is usually a bad idea if you are looking for a specific result.
• Be specific: Describe exactly what you want the AI ​​to do, without assuming it will interpret ambiguous language the same as you do. Includes details and examples, to make sure that you and the AI ​​are ‘thinking’ the same thing and thus ensure that the result matches what you were asking for. This is best explained -of course- in the form of an example.

The following image is generated from the instruction “A rabbit as Harry Potter” (‘a rabbit like Harry Potter’). We are giving an example to the AI ​​(Harry Potter) and at the same time using simple language (X as Y).

• Use positive language: describe what you want to get, not what you don’t want.
• use adjectivesif possible several (symmetrical, abstract, realistic, complex, classic, colourful…)
• Mention the quality what you want to get, so the AI ​​doesn’t wander: things like ‘high quality’ or ‘8K’ can help.
• Learn Art History. Above all, contemporary art: mention authors (painters, illustrators…) or styles (things like ‘Japanese poster graphics’, ‘synthwave’ or ‘1990s point and clicks 16bit adventure game’) to define the result.
• Be aware of biases: AI models are powered by content from the internet, so you have to think how to counter them with your instructions.

My last advice: be very careful with literally copying and pasting prompts without looking at what AI they are intended for, because each one is a world. For example, the ‘prompt’ “cat::1 eating ice-cream::3 beach::1” should generate an image in which these three elements appear (cat, eat ice-cream, and beach), but in which the The greater weight of the image falls on ‘eating ice cream’… as long as we are using MidJourney.

But look what happens when we use it in DALL-E 2:

If DALL-E 2 reads ‘cat’, it will focus on the cat. They don’t care about priority instructions.

For ChatGPT

The basic tips from the previous list are maintained, of course: yes to simplicity of language, no to ambiguity, provide details when needed etc. But the truth is that the conversational capacity of ChatGPT allows us to address this AI in a completely different way than image generators and, in fact, its versatility sometimes forces us to elaborate long and complex ‘prompts’ to achieve the desired result.

We have already seen in these pages how we can trick ChatGPT into bypassing its own content policies (instructing it to play an RPG) or to behave like a virtualized Linux console.

As you will see by reading each link, the trick with ChatGPT lies in giving it clear instructions on how we want him to react to what we write (remember that, within the same conversation, ChatGPT keeps ‘memory’ of previously entered instructions).

So if we want generate a ‘table of contents’we can enter something like this:

“You are in the world debate championship. The debate motion is “This assembly supports universal income for all citizens.” The government coalition is in favor of the motion. Think of the arguments in favor of the motion, using a hierarchical table of contents that categorizes the arguments into at least 3 topics, then write the coalition speech. Then do the same for the opposition, who should argue against the coalition arguments.”

And if what we want is for ChatGPT to act as a translator that not only translates but also improves the style of the entered text We can resort to this other prompt, extracted from an interesting repository of ChatGPT commands:

“I want you to act as an English translator, spell checker and improver. I will speak to you in any language and you will detect the language, translate it and reply in the corrected and improved version of my text, in English. I want you to replace my simplified words and sentences of A0 level with more beautiful and elegant higher level English words and sentences. Keep the same meaning, but make them more literary. I want you to only answer the correction, improvements and nothing else, don’t write explanations. My first sentence is [insertar]”.