Prompt Engineering via Prompt Patterns — Outline Expander Pattern
The article is part of series: Prompt Engineering via Prompt Patterns
You can switch to video version of this article
When we talk about content generation, we are usually talking about stories, movie or tv series plots, articles, books, research papers and what not. Sometimes we just open up our favorite editing software and start typing, but most people take a more formal approach, think through the entire ‘content’ from start to finish. Note I am going to use the word content to mean all the examples of content I mentioned earlier.
More often than not, such content has a pre defined structure or pattern to it. Like most stories would start out by introducing the environment or setting followed by introducing the characters, then a problem, then the struggle followed by climax. If it is a tv series, it would break down these steps into multiple episodes or seasons, where initial ones would introduce the characters and then following episodes would introduce a problem and struggle where each episode can be fully independent or sometimes a global theme or a super story flows from starting season/episode to the end with filler episodes primarily being semi independent of overall plot, but nudging the overall theme forward. I just took a story or season example, but take any kind of content, and it helps to define an overall theme or structure for the overall content, sort of define an outline and then fill each of the parts individually with content.
By outline I mean a sort of hierarchy. If you have used Microsoft word for instance, it allows you to generate a table by using heading types as hierarchy. Your html document has H1, H2 headings to let you break down the content into a hierarchy. You can generate an outline for almost any content of any type. And if that is true for most content, the converse should also be true that you should be able to generate any content starting with a hierarchy, or outline. This lets you be in total control of flow of content from start to finish, and gives your readers a much more engaging experience right until the end. Anyone who enjoys watching tv series must have had a rub with horrible and distasteful endings, primarily because the authors never imagined their series to run for the number of seasons it did, or the scare of series being cancelled being on their mind, so the final seasons or episodes are aired much after the expiry of original outline. I just want to highlight how important original outline of any content is before jumping to our pattern.
So right on cue, lets introduce the Outline Expansion pattern. We already covered the main idea so no surprises here. To generate any content, we start by picking up a topic, and then then generate a breakdown of that topic into headings or bullet points like coming up with a title of a book, followed by names for each chapter in the book. We can continue expanding further but let’s just stop ourselves here for a second. The concept of this pattern is to give you tools to write content by focusing on one part of the outline at any given time. You start with topic, then you concentrate on generating subheadings or constituent parts for that specific topic, then you pick one specific subtopic and repeat the whole thing or just generate the contents for that subtopic. Its kind of like growing a tree, where on each decision point, you decide whether to grow a branch, or a leaf and go no further. The interesting bit is that you are spared most of this effort by large language models like ChatGPT as they are able to generate subtopics for almost any topic you provide them, until of course you are working on something really new. They can even generate subtopics for each bullet point and content for these for you.
To use this pattern, you need to use a prompt like below, and note how multiple prompt patterns are used here to make up the prompt below
Act as an outline expander
Generate a bullet point outline based on the input that I give you, and then ask me for which bullet point you should expand on
Create a new outline for the bullet point that I select
At the end, ask me for which bullet point to expand next
Ask me for what to outline
This is it, we used persona pattern to tell ChatGPT that it is an outline expander which configures it to work with outlines where possible, and it would neatly and implicitly try to use outlines from now on. Then we ask it to generate a bullet point outline based on the topic you provide, say our solar system and it should neatly come up with 8 or 9 bullet points numbered 1–9. Note that it was a closed and predictable example topic. But for other topics, say animals found in XYZ country, the list could be potentially endless if not categorized correctly, so you should be able to intervene with explicit instructions like each bullet can have at most 3–5 sub bullets. This is followed by instruction to ask you which bullet point to expand. This is meant to stop the model from being over efficient and start generating content for first item. You need to be in control of what it does.
Next you instruct it to create a new outline for the bullet point that you select. Now you can specify the bullet point by number, and then it is smart enough to generate sub bullets in the format 1.3 or 3.5 to make it easier for you to specify which bullet point anywhere in the outline you are referring to. Note you are not constrained to use the numbering system ChatGPT starts using. In the initial prompt, you can introduce instructions to use a particular numbering system like the bullets should be numbered using the pattern followed by the pattern like roman numbers, alphabets etc.
Next you tell it to generate outline for the bullet point that you select and in the end asking you which bullet point to generate content for. Now in this example prompt, you are defining the entire outline before you would start to generate content, so it would expect you to keep prompting you with which bullet point to expand next, until you respond with generate content or write text for bullet 1.3.2.
You can add a further instruction to the prompt to be able to generate content for any bullet point in the outline easily. You can add an instruction like Whenever I type write
Note the BulletPoint placeholder is referring to the bullet point number like 3.1.2 or something, while paragraphs is the count of paragraphs to generate. Again this is a guideline and you can add further instructions using format like above to add further instructions to the model to configure the model as per your desire.
Its techniques like these which can be used to configure the model in a host of situations, not specific to our outline expander pattern. The entire purpose of our prompt patterns and prompt engineering series is to equip you with tools to configure large language models in the most creative ways possible. You should consider these techniques as examples to use in your unique situations which we obviously can’t even begin to cover. Important point is, you should strive to never write a prompt the novice way again once you have learnt patterns like this one.
Coming back to outline expander pattern, note though that the list of topics or subtopics it generates might not be what you expected, and you might need to instruct it to modify it according to your expectations iteratively. You can do this by asking it to add further topics you provide to bullet 3.2 or instruct it to classify the list of topics differently like for the animals examples I used, you can instruct it to classify them using biological families or something. Similarly for the planets example, you can instruct it to add number of moons to your subtopics list for each planet.
Another issue to keep in mind is that there is a memory limit of large language models and after sufficient content generation, if you ask it to create outline for say 1.2 after working on 1.x for a while, it might have forgotten about it. Faced with such a rare situation, you still have an option. First off you can watch the memory and context related videos on our channel to be aware of what I mean by memory limits of LLM. Secondly, you need to handle it just like talking with a human. You need to remind the model about the context. So if you were copy pasting the outline to a word editor, paste the relevant list of topics back to the model via prompt. And to be able to do that, keep the outline separate from content in your editor. Once ‘reminded’ with the relevant section of outline, the model would start following your instructions like before.
Few minor potential issues aside, Isn’t this pattern cool? You are in full control of your content generation and can copy paste generated content into your favorite word editor after you have your ‘table of content’ defined first. Starting with a satellite view, to bird’s eye view to immersive experience. Systematic, coherent, and controlled flow of content achieved with such a simple prompt. Level up your expectations as once this article goes viral, bad content would be a thing of the past.
I should better get out of my overly optimistic day dream and let you off to focus on writing amazing content using outline expander pattern 😊 If you liked the article, please clap and share. You can consider subscribing to our YouTube channel as well. Thank you!!!.
