Best Practices For Creating with AI

Hello, Curious Writers,

I recently developed a guide for my students on best practices when using AI. AI is an emerging technology, and many people are curious about its use but need help knowing where to start. My students and several colleagues found this information helpful, so I am sharing it here.

Writing with AI

For years, scientists tried to develop a language model by training AI to know and follow the English language rules (much like you learned in English class). However, the amount of knowledge and nuance needed to understand natural language made the task impossible.

Current AI works off of a pattern recognition scheme. It has been trained with large datasets of text to recognize patterns, identify relationships between words, and anticipate grammatical structures and sentences. It predicts the likelihood of a sequence of words. This is similar to when your phone suggests words based on your text message.

This means, as good as it is, it is not crafting knowledge for you. It is simply returning words based on themes and patterns. This is why you can input the same information and get different responses. 

Humans must bring wisdom to the situation and critically assess the quality of the knowledge being returned. We must continue to contextualize, interpret, and critically evaluate AI-provided information to ensure its relevance and accuracy.

AI Threads

In Chat-GPT, treat every thread like a new person or assistant that you need to train to do a task. If you haven’t asked it to do that task in a while, you will need to retrain it (as it can forget—it doesn’t store data). Also, at times, the thread will have a personality or go down a rabbit hole. You either need to retrain it or start over with a new thread.

Interpreting Research and Interviews with AI

As part of a class experiment to test the capabilities of AI to write an article, we gathered factual information and conducted interviews. In our test, we found that it is best to give AI elements of the information that you find to be most relevant.

• AI did not do well when asked to incorporate quotes from the raw transcript.
• It made up quotes when asked to summarize and pick five great quotes from a transcript.
• When given a summary of the interview and five quotes from a human-curated list, it produced the best topic.

While AI is great at finding patterns and synthesizing data, when asked to write a report or article, it produces its best work when you give it only the information you find to be relevant. Too much information produces a poor product.

Know AI’s limits.

What it is good at.

• Leverage AI for brainstorming and first-draft content generation.
• Use AI to analyze the writing style and tone (active voice).
• Giving recommendations on changes.
• Seeing patterns and outlining content.

What it is bad at.

• Making changes. It tends to want to change everything, like using a thesaurus.
• AI tends to hallucinate.
• AI does not have access to up-to-date information unless plugged into search engines.
• Fact-checking.

AI Prompts

AI is only as good as the prompt.

• Be as specific as possible with the ‘specifics’ and ‘do not’ sections to avoid any potential misunderstandings.
• When mentioning a style, provide an example or reference, as this will give the AI a concrete point of reference.
• Be mindful of the instructions’ complexity. If the prompt becomes too long or convoluted, the AI might not be able to follow all parts equally well.
• Encourage interrogation and iteration, meaning the user should ask AI for feedback and feel free to refine and rephrase the prompt based on the AI’s feedback or initial responses.

A good prompt format:

• Action Plan: “In a minute, I am going to [what you are going to ask it to do].” This sets up an immediate intention for the AI to anticipate.
• Role-Playing Instruction: “I want you to act as if [person, process, or object].” This provides context and a point of view for the AI’s response.
• Desired Format: “Create [format you want].” Specifies the structure the user wants, such as an essay, list, dialogue, etc.
• Content Specifics: “That contains [specifics].” This ensures the AI includes all the necessary details.
• Specific Instructions: “Make sure to [what you want it to do].” These are additional actions or elements that the user wants to be emphasized.
• Limitations: “Please do not [what you do not want it to do].” Defines boundaries and constraints for the AI’s response.
• Stylistic Preferences: “In the style of [specific reference or example].” It gives the AI a model to emulate in its response.
• Tone and Audience: “Make sure to use a [tone] for [audience].” This specifies the mood of the response and the target audience, which is crucial for tailoring the language and approach.
• Feedback Request: “What else do you need to know to complete this? What am I missing?” This invites the AI to ask clarifying questions if the initial prompt isn’t enough to start with.

Example:

In a minute, I am going to [what you are going to ask it to do]. I want you to act as if [person, process, or object] and create [format you want] that contains [specifics]. Make sure to [what you want it to do]. Please do not [what you do not want it to do]. In the style of [ specific reference or example]. Make sure to use a [tone] for [audience]. What else do you need to know to complete this?

Illustrating with AI

Midjourney Prompts:

• Word choice matters. Be specific and use synonyms for more nuance. For example, use terms like “tiny” or “gigantic” instead of “small” or “large.”
• Detail is key when dealing with plural words. Specify exact quantities—say “three dogs” rather than “dogs.” For groups, explicit terms like “a cluster of stars” can be more effective than “stars.”
• Concentrate on desired elements rather than undesired. If you wish to exclude an item, use direct exclusion parameters like “–no” for better results.
• The length and detail of prompts vary. But, to steer the creative direction, integrate crucial elements into your prompt. A more detailed prompt affords greater precision but can limit creative freedom.

Consider the following for richer prompts:

• Subject: Specify whether it’s a being, a place, an item, etc. (person, animal, character, location, object)
• Medium: Define if you’re seeking a photograph, a sketch, a 3D model, etc. (photo, painting, illustration, sculpture, doodle, tapestry)
• Setting: Are you imagining a scene set indoors, in the wild, or perhaps in space? (indoors, outdoors, on the moon, underwater, in the city)
• Lighting: Describe the lighting—should it be dim, harsh, or maybe twilight? (soft, ambient, overcast, neon, studio lights)
• Color: Choose from a spectrum of tones—bold, subdued, or perhaps a specific color theme. (vibrant, muted, bright, monochromatic, colorful, black and white)
• Mood: Decide on the vibe, whether tranquil, lively, or somewhere between. (sedate, calm, energetic)
• Composition: Determine the viewpoint—should it be a panoramic shot, a detailed close-up, or an aerial perspective? (portrait, headshot, close-up, birds-eye view)

DALL-E Prompts:

A good prompt consists of two parts: Content and Modifier.

DALL-E uses prompt modifiers to influence various aspects of the image, style, and quality. According to research by Jonas Oppenlaender, there are six types of modifiers.

• Subject Terms: Wording that describes the main subject of the image (e.g., “a landscape at dawn”)
• Style Modifiers: References that direct the AI to create images in a particular artistic style (e.g., “in the animation style of Pixar”)
• Image Prompts: URLs to images or downloaded image files that provide further context for image creation
• Quality Boosters: Wording to enhance the aesthetic qualities and details of images (e.g., “award-winning,” “intricate,” or “stunning”)
• Repetitions: Repeated subject or style terms to increase the focus DALL-E places on a particular part of the image (e.g., “a very, very, very, very, very beautiful landscape”)
• Magic Terms: Terms and phrases that inject randomness and creativity (e.g., “control the soul”)

Example:

“An ultra-modern kitchen showcasing a sleek, stainless steel coffee machine, rendered in great great great detail, with a style reminiscent of modern architectural photography that would trend on design blogs.”

This prompt includes content (“modern kitchen,” “coffee machine”), style modifiers (“reminiscent of modern architectural photography”), quality boosters (“trending on design blogs”), and repetition (“great great great detail”) to produce a graphic that would otherwise take hours to create.

Tips:

1. Know Visual Art Language: (style: impressionism, aspects: hue, saturation, location: foreground).
2. Balance Creativity with Clarity: descriptive language helps, but too much can lead to unexpected results.
3. Limit Text Rendering: AI image generators can handle short text, but it still needs to be completely accurate.
4. Stay Away from Negative Prompting: These generators work on creating visual aspects that align with what you want. Telling it what you do not want does not work either.
5. Be Specific and Detailed: The more specific your prompt, the better the image quality. Include details like the setting, objects, colors, mood, and any specific elements you want in the image.
6. Mood and Atmosphere: Describe the mood or atmosphere you want to convey. Words like “serene,” “chaotic,” “mystical,” or “futuristic” can help the AI set the right tone.
7. Use Descriptive Adjectives: Adjectives help in refining the image. For example, instead of saying “a dog,” say “a fluffy, small, brown dog.”
8. Consider Perspective and Composition: Mention whether you want a close-up, a wide shot, a bird’ s-eye view, or a specific angle. This helps frame the scene correctly.
9. Specify Lighting and Time of Day: Lighting can dramatically change the mood of an image. Specify if it’s day or night, sunny or cloudy, or if there’s a specific light source like candlelight or neon lights.
10. Incorporate Action, Movement, or Emotion: If you want a dynamic image, describe actions or movements. For instance, “a cat jumping over a fence” is more dynamic than just “a cat.”
11. Use Analogies or Comparisons: Sometimes it helps to compare what you want with something well-known, like “in the style of Van Gogh” or “resembling a scene from a fantasy novel.”
12. Specify Desired Styles or Themes: If you have a particular artistic style or theme in mind, mention it. For example, “cyberpunk,” “art deco,” or “minimalist.”
13. Iterative Approach: Sometimes, you may not get the perfect image on the first try. Use the results to refine your prompt and try again.

I hope this helps you when you are playing with generative AI. Happy creating!