Simple legal AI prompting tip: 🥵 Hashtags are hot again

Simple legal AI prompting tip: 🥵 Hashtags are hot again

Cecilia Ziniti

Jun 6, 2024

🥵 Simple legal AI prompting tip: use hashtags. Yes! Hashtags are hot again. Hear me out. 

# Hashtags are how you can give AI instructions it can understand. 

To an AI, consecutive hashtag symbols (#, ##, ###)  represent headers. Like this:

# Header 1
## Header 2
### Header 3

And so on. 

The benefit? In this way, hashtags allow you to give the AI a beautiful, sectioned with headers memo of what you want. Kind of like how in a court brief or a contract, the headers guide the reader or the implementer of your document with what you want. They matter! They're better than unstructured, stream of consciousness prose. 

AI is like a person - a very capable intern who has passed the bar exam. So just like it's easier to get people to do things for you when they can understand you, speaking AI's language immediately boosts its performance on tasks you assign it. 

🧑‍💻 PS - Fun fact. The reason AI understands headers for hashtags (and *asterisks* for bold, _underlines_ for italic, | pipe characters for tables |  and other seemingly weird characters) is because of … open source.

OpenAI’s models were trained on a lot of publicly available software code because code uses language to show logical operations and connections in a structured way, enabling AI to simulate reasoning. So AI models trained on software code perform better than those trained on just words. 

The result? AI models like to read documents formatted the way comments and files accompanying open source code are formatted in places like Github.

Enter markdown language, a form of HTML. AI reads it well. And in Markdown, headings are denoted by… you guessed it… hashtags! So using Markdown formatting in your prompts will help the AI better understand your request.

Use it to delimit the different parts of your prompt. Like in the example below from one of GC AI’s official prompts:

# Task

Please review the legal terms for a SaaS product we are considering buying. I'd like you to provide a table of the top 5 issues that you might highlight in these terms as risky or unusual for customers of this product that I should be aware of. I am most interested in whether and how this service can use my company's data for AI model training or machine learning. 

## Opening Paragraph

Give me a general assessment of whether this vendor has thought about the use of AI and data. Does the contract show they are concerned about this with their customers?

## Table to be provided after the paragraph

The columns should be: item, why it's risky, the exact quote in the terms of use that you are relying on for that assessment, and then an emoji representing if it's something I should further review (🟩 for review not needed; 🟨 for consider reviewing; 🟥 for this must be reviewed by a lawyer).

## Link to terms

And since, like an overeager intern, AI likes to please you, it will respond in kind by outputting Markdown format, which is what allows it to create pretty tables like the one in the screenshot from GC AI below.