In a nutshell, prompting any AI is actually being able to talk to and instruct a computer using natural language. As this evolves, all the syntax and engineering will be less and less important and instead will get close to how we talk to each other. Although for now, you can leverage exponentially more potential when using a wide range of prompt engineering techniques. Experts in various fields of every kind almost have a language of their own. The same goes for programming languages, data science and every other topic. Depending on the specific model and it's training, the large models will have nearly everything to some degree.
Chat/Text Prompting:
- Absolute Value Output: After lots of prompting there are patterns that begin to take shape. One of these is the amount of instructional prompting that can be achieved per-prompt. Although depending on the specific task, there does seem to be a sort of minimum to maximum amount of task achievement per-prompt. When you attempt to go far over this, the model either fills in the gaps or gives incomplete answers. Get a feel for the amount of task each prompt can achieve and work around this. Take your bigger tasks and break it down into simple achievable tasks.
- Chain Of Thought: This is a known technique by a researcher and employs an effective strategy of instructing the LLM to break down the task into smaller tasks and work though the problem sequentially similar to how we work out math problems.
- Programming: Although this will be less and less useful, if you are already familiar with programming, techniques used in programming seem very effective for prompting. Use bullet-points, lists, outlines and all kinds of other familiar formats. Since AI currently(these will likely be trained more properly in the future) is from random scraped web data, think of the most common articles and papers.
Image Prompting:
- Description: Models are trained from images that have been tagged with descriptive keywords. Remember all those hashtags and keywords we labeled our images and posts with? Well that formed the basis for image prompting.
- Example: nice sunset on a beach, vibrant yellow red and orange colors,
- Keywords: You can enter phrases or even sentences and those can work better in certain cases. Basically get as precise and direct to the point as possible. Descriptive keywords can be one or a few words. Use commas to arrange your prompts.
- Example: nice sunset on a beach, vibrant yellow red and orange colors, artist painting, high contrast, 4k
- Strengths: There is syntax to increase the strength of specific words or phrases within your image prompt. Place words inside parenthases () and you can also stack for more strength.
- Example: (nice sunset on a beach), vibrant yellow red and orange colors, ((artist painting)), high contrast, 4k
- Negatives: Since image generation is scraped from large unsupervised datasets, it has not been filtered from all kinds of bad quality and undesirable images. Negative prompts are needed to remedy this. The same general concepts apply to the positive prompts but it's just instructed to avoid everything in the negative prompt.
- Balancing: Ideally the positive and negative prompts should be somewhat balanced in terms of text length and the 'measure of demand' of the concepts included. There are padding settings to automatically fill in prompts to balance them depending on the tool used but may be good to keep in mind.