Tell the LLM the business context

Employees do better when they have more business context. The same is true of LLMs! To do its best work, the LLM needs to know why it’s being prompted, where its input came from, how its output will be used, and how its output will be judged. Many prompters try to tell the LLM how to achieve some task, but it’s often better to just give it the business context.

An example. Your cooking blog has a very plain homepage. Wouldn’t it be nicer, you think, if each link to a blog post included an intro paragraph to draw readers in? And isn’t this the kind of thing an LLM should be great at writing? So you write your first prompt:

Summarize the following article.

You run it on your blog posts, but the responses you get are inconsistent and mediocre. You pile more instructions into the prompt, until you end up with:

Summarize the following article.
No more than 2-3 sentences.
Make it engaging so the reader wants to know more.
Never refer to "the post"; instead summarize directly.
Never refer to "the author"; instead use "I".
Here are some examples: ...

But the LLM rarely remembers all the rules. Why is it so stupid, you wonder?

Instead, try telling the LLM the business context. Here’s an alternative prompt: is Caleb Smith's blog about cooking.
Each post link on the homepage has the post title, plus a teaser paragraph.
You are given a post's title and content, and you write its teaser paragraph.
The goal is to convince readers to click.
Here's one example: ...

Imagine you’re describing this task to a contractor. You wouldn’t tell them how many sentences to use; you’d just tell them the business context. They’ll figure out what’s appropriate.

The problem with “Summarize the following article” is that the LLM starts out in a superposition of all the business contexts in which it might have been asked for a summary, such as:

As the LLM produces output, it collapses into one of these contexts. So some “summaries” will look like critiques; others will look like keyword lists.

I have seen lots of LLM prompts that avoid providing the business context. One reason, I think, is that programmers treat the prompting as programming in English. Programmers are used to describing the how, but not the why. And programmers love modularity, where code is reusable and modules don’t know about each other. But the “business context” principle is anti-modular, because then the summarization module has knowledge about the whole app and business.

Tagged #llm, #machine-learning, #programming.
👋 I'm Jim, a full-stack product engineer. Want to build an amazing product and a profitable business? Read more about me or Get in touch!

More by Jim

This page copyright James Fisher 2024. Content is not associated with my employer. Found an error? Edit this page.